chore(release): 1.0.0b1 — the v1 target, staged as a beta

All seven v1 capabilities are landed (ROADMAP's v1 target is met), so this is
the first release of the 1.x train. Staged as a beta rather than cut final on
the operator's call: per the canonical policy `-beta.N` means feature-complete,
external testing, no new features, focus is on bugs — which is exactly this
state, with a cross-frontier bug-hunt panel outstanding on U7's diff.

The repo learned this sequencing the hard way once: v0.2.0 was tagged and
announced while a contract panel was in flight, the panel found three defects in
the code just released, and v0.2.1 shipped within the hour. A beta is the
designed answer to that, not a workaround for it.

ALSO FIXES A SECOND COPY OF THE VERSION, found while cutting this one.
`booth.__version__` was the literal `0.1.0` and had been wrong through six
releases. It is now read from pyproject.toml — deliberately NOT from
importlib.metadata, which describes a different artifact: this repo has no build
step and no install step (booth.service runs uvicorn with WorkingDirectory set
to the tree), and the venv was carrying a vestigial booth-0.3.0.dist-info with
no package directory behind it. Installed metadata therefore reported 0.3.0 for
a tree at 1.0.0b1 — confidently wrong and varying by environment, which is worse
than a literal that at least fails the same way everywhere.

booth/__init__.py is also, it turns out, effectively stdlib-only: scripts/booth
imports booth.links / booth.marks / booth.manifest under the system python3 with
no venv, and every one of those executes the package root first. Nothing
asserted it. test_stdlib_only now covers __init__, and the no-venv import path
is verified under python3.11 reporting 1.0.0b1.

642 tests green.

CLAUDE.md

@@ -62,8 +62,9 @@ test is the only thing standing here.
 No database. `ls ~/booth-data` tells you everything the service knows.
 
 Per-booth operator state is a **dotfile inside the booth**: `.forever` (keep),
-`.blurred` (one rel per line), `.marks.json` + `.marks.lock` (judgment), `.pins`
+`.viewed` (last deliberate look — U4's "viewing is activity"), `.blurred` (one
-(link-board pin ids), `.uploaded` (upload-booth marker). `booth_items()` skips `name.startswith(".")`, so a new
+rel per line), `.marks.json` + `.marks.lock` (judgment), `.pins` (link-board pin
+ids), `.uploaded` (upload-booth marker). `booth_items()` skips `name.startswith(".")`, so a new
 dotfile costs nothing in item counts, galleries or zips. That skip is why the
 dotfile is the right shape for new operator state — use it rather than
 inventing a sidecar-per-item.
@@ -84,6 +85,12 @@ doing so, never resolved the caption — the operator's "zoomed images lose
 their annotations" bug. It was not a rendering bug; it was three readers of one
 truth.
 
+**U3 extended this to the verbatim path.** A booth's own `index.html` now gets
+its chrome from `/_booth/embed.js`, which *places* server-rendered fragments and
+never builds one. The fragments come from the same `_ask_inline.html` macros the
+gallery page uses, handed over `/b/<name>/embed.json`. A second renderer in
+JavaScript would be the same bug in a new language.
+
 ### 4. Re-export, don't move-and-break
 
 Names that moved from `app.py` to `items.py` (`classify`, `doc_kind`,
@@ -121,9 +128,14 @@ the operator's judgment being quietly misfiled.
 
 Current rules: items `sorted(rel)`; the zoom ring is that order filtered to
 images; captions resolve over a sorted scan; marks `(created, id)`; legacy
-import `(mtime, name)`; link rows pinned-then-newest. `ROADMAP.md` carries the
+import `(mtime, name)`; link rows pinned-then-newest; a verbatim report's embed
-table and the two places still undecided (U7 sections and compare pairing, U6
+anchors in document order, its tail in payload order, its questions in
-bench listing).
+declaration order. `ROADMAP.md` carries the table and the two places still
+undecided (U7 sections and compare pairing, U6 bench listing).
+
+U3's rows are the first that bind **across a language boundary** — decided in
+Python, honoured in JavaScript. A string assertion cannot see that, which is
+why `tests/test_embed_browser.py` exists.
 
 When you add an ordered surface, state its rule in the docstring. If you cannot
 state it in one line, it does not have one.
@@ -178,8 +190,15 @@ one caused an outage.
    is that template work needs a restart to see, and that price is the point.
    `test_templates_do_not_hot_reload_from_disk` holds the line.
 
-**So: after ANY edit here — Python or template — the live service is stale until
+3. **`booth/static/embed.js` is the third thing that would have hot-reloaded,
-you restart it.** If you are touching this repo while the operator may be using
+   and it does not.** U3 gave the service a static asset living in the
+   deployment root; it is read ONCE in `create_app` and served from memory with
+   an ETag over its content, for exactly the reason above. Same rule, same test
+   shape (`test_embed_js_does_not_hot_reload_from_disk`). Anything else this
+   repo learns to serve from disk inherits the rule — read it at startup.
+
+**So: after ANY edit here — Python, template or static asset — the live service
+is stale until you restart it.** If you are touching this repo while the operator may be using
 the service, either restart promptly or expect him to be looking at the old
 version. Never leave the tree in a state where a restart would 500.
 
@@ -196,6 +215,15 @@ curl -s localhost:8090/healthz         # the live service (systemd --user)
 systemctl --user restart booth.service # after a code change, to see it live
 ```
 
+`tests/test_embed_browser.py` drives a real Chromium against a real uvicorn on
+an ephemeral port — the only place U3's placement and `form=` binding can be
+observed at all. Browsers are NOT downloaded per project; they live box-wide in
+`/opt/ms-playwright`. The file **skips rather than fails** when playwright or a
+usable browser is missing, so the suite stays green anywhere. If those tests
+start skipping on this box, the pinned `playwright>=1.60,<1.63` in
+`pyproject.toml` has drifted past the shared store — read the comment there
+before raising the bound.
+
 `booth.service` is a user unit installed to `~/.config/systemd/user/`. The repo
 copy is the source; edits there need a `daemon-reload`.
 

README.md

@@ -15,16 +15,18 @@ filesystem *is* the state.
 - **Live:** http://10.100.10.50:8090/ (nh3-dev) · linked from Homepage → *Apps → The Booth*
 - **Data dir:** `~/booth-data/` on nh3-dev (one subfolder per booth)
 - **TTL:** 24h, measured from the newest mtime in a booth's tree (it lives while
-  you're touching it, self-destructs 24h after you stop)
+  you're touching it, self-destructs 24h after you stop). Two things hold a booth
+  open past that: the `.forever` sentinel, and **an unanswered question** — see
+  *Lifetime* below. **Opening a booth page is activity**; polling it is not.
 
 ## How a session posts
 
 A booth is **just a folder** under the data dir. Three ways, cheapest first:
 
 ```bash
-# 1. On nh3-dev — the helper (services/booth/scripts/booth):
+# 1. On nh3-dev — the helper (scripts/booth):
-booth add my-run  out/a.png out/b.png       # creates booth + copies, prints URL
+booth add my-run out/a.png out/b.png --why "pick the denoiser, v3 on the left"
-booth new  my-run                            # empty booth, then cp/mv into ~/booth-data/my-run/
+booth new  my-run --why "..."                # empty booth, then cp/mv into ~/booth-data/my-run/
 booth url  my-run                            # just print the URL
 booth ls                                     # list booths
 booth rm   my-run                            # wipe now (TTL would anyway)
@@ -39,6 +41,25 @@ rsync -a ./out/  nh3-dev:booth-data/my-run/
 
 Then hand the operator `http://10.100.10.50:8090/b/my-run/`.
 
+### Say what it is — `--why`
+
+**`--why` is one line telling the operator what he is looking at and why.** It
+lands on the index card and on the booth page next to your handle (taken from
+`$ALTHING_HANDLE`), stored as `.booth.json` in the booth.
+
+It is optional and nothing breaks without it — a booth with no announcement
+renders as `unannounced`, which is also what every booth created by `rsync` or
+a bare `mkdir` looks like. But a booth that cannot say what it is has no way to
+ask for attention except by posting its URL somewhere else, and that is exactly
+how the link board ended up 69% dead rows. **The booth is the place to say it.**
+
+```bash
+booth add r18-ab out/*.png --why "which denoiser — v3 left, v4 right" --title "R18 A/B"
+```
+
+A second `new` or `add` on the same booth updates the why and keeps the
+original creation stamp: the booth appeared once.
+
 ## Checking that controls can actually be clicked
 
 ```bash
@@ -140,7 +161,46 @@ is exactly why the direct `×` was worth adding.
 an ephemeral booth could only be kept from a shell. The `/keep` route and the
 CLI verb both already existed; only the button was missing.
 
-## Kept boards — the one exception to the 24h rule
+## Lifetime — derived, not declared
+
+A booth is in exactly one of three states, and only the first is a button you
+press:
+
+| state | what puts it there | swept? |
+|---|---|---|
+| **kept** | you pressed `keep` / dropped `.forever` | never |
+| **held** | an **unanswered pick**, or a `.marks.json` the service cannot read | not while that holds |
+| **ephemeral** | everything else | 24h after the last activity |
+
+**An open question holds its own booth.** A session that runs `booth ask` does
+not also need to `keep` the booth — the booth cannot be swept while the operator
+still owes it an answer, and it is released automatically when he answers. A
+*partially* answered multi-question pick still counts as open, so a review in
+flight is never swept out from under him. The index card and the booth header
+say `held until answered` where the countdown would be, so a booth that has
+stopped counting down always tells you why.
+
+**Viewing is activity.** A deliberate GET of a booth's own page — the gallery, a
+verbatim report, the zoom view, the marks page, a zip download — resets the
+clock. If the operator is still looking at it, it is still alive. Browsing the
+index does **not** count, and neither does a session polling `marks.json` or
+`booth marks --wait`: machine reads are deliberately excluded, so an agent
+cannot hold its own booth open by waiting on it.
+
+**A held booth is still yours to delete.** The hold is protection from the
+timer, never from you: `booth rm`, the UI ×, and `DELETE /b/<name>` all work
+exactly as before. `sweep_once` is the only thing that honours a hold, exactly
+as it is the only thing that honours `.forever`.
+
+**Why this exists:** `.forever` used to be the only way to say three different
+things — "this is durable", "I haven't answered yet", and "I'm still looking at
+it" — and the measurement showed it carrying all three. On 2026-09-22, 17 of 24
+live booths (70%) held the sentinel, up from 54% the day before; three of the
+four booths in the fleet awaiting an answer had been pinned by hand as well.
+Only the first meaning is what `keep` means. The other two the service already
+knew and did not consult.
+
+### Kept boards — the explicit pin
 
 A booth containing a **`.forever`** dotfile is **never swept**, and renders in
 its own **Kept** lane at the top of the index (blue top edge, `★ kept` badge, no
@@ -149,6 +209,7 @@ still ephemeral, so nobody inherits a cleanup chore they didn't ask for.
 
 ```bash
 booth keep   my-board      # drop the sentinel — exempt from the sweep, forever
+                           # (NOT for "waiting on an answer" — the pick holds it)
 booth unkeep my-board      # release the pin — the board rejoins the sweep
 booth rm     my-board      # delete it NOW (works on kept boards; says so when it was kept)
 
@@ -419,11 +480,10 @@ wipe it from there. Release is reversible — press keep again and nothing was
 lost. From the CLI, `booth rm <name>` deletes a kept board immediately and
 tells you it was kept.
 
-**Do not "unkeep and let it expire."** Removing the sentinel *bumps the booth
+**Do not "unkeep and let it expire."** **Releasing a board is activity** — you
-directory's mtime*, and a booth's age is the newest mtime in its tree — so a
+just touched it — so a released board's clock **resets** and it survives another
-released board's clock **resets** and it survives another full TTL.
+full TTL. Unkeep-and-wait is a 24-hour delay, not a delete. Use the × or
-Unkeep-and-wait is a 24-hour delay, not a delete. Use the × or `booth rm` when
+`booth rm` when you mean now.
-you mean now.
 
 ## Ops
 

ROADMAP.md

@@ -1,7 +1,12 @@
 # The Booth — roadmap
 
 Design: [`docs/design/information-architecture.md`](docs/design/information-architecture.md).
-Current version: `0.2.1` (U1 + U2 landed; extracted from eshpfi 2026-09-21).
+Current version: `1.0.0b1` (**U1 through U7 landed — every v1 capability is
+in**; extracted from eshpfi 2026-09-21). **The v1 target is MET and staged as a
+beta** (operator, 2026-09-22): feature-complete, external testing, no new
+features — the remaining work is bugs. `1.0.0` final is cut when the beta
+survives; per the canonical policy an rc would be cut from the same commit,
+but a beta may still take fixes.
 
 ## v1 target
 
@@ -12,18 +17,52 @@ defect — not a wish. The measurements are in the IA doc.
 |---|---|---|---|
 | 1 | ~~**One item record**~~ — **landed `ce598b3`** | captions never reach the zoom view (never sent, not lost) | U1 |
 | 2 | ~~**Marks**~~ — **landed `c7f9437`, released `v0.2.0`** | 5 mechanisms for 1 job; operator→session loop runs through chat | U2 |
-| 3 | **Declared embed seam** — `/_booth/embed.js`, chrome mounts via DOM | 6 regexes injected into arbitrary author HTML, load-bearing for asks | U3 |
+| 3 | ~~**Declared embed seam**~~ — **landed `87e2c53`, released `v0.5.0`** | 6 regexes injected into arbitrary author HTML, load-bearing for asks | U3 |
-| 4 | **Derived lifetime** — open marks pin; viewing is activity | 54% of booths on the `.forever` escape hatch | U4 |
+| 4 | ~~**Derived lifetime**~~ — **landed `c3a97c1`, released `v0.4.0`** | 70% of booths on the `.forever` escape hatch (54% when first counted) | U4 |
-| 5 | **Self-announcing booths** — `.booth.json`, provenance on the index | job 5 had no home, so it lived on the link board as 145 dead rows | U5 |
+| 5 | ~~**Self-announcing booths**~~ — **landed `c015a91`, released `v0.3.0`** | job 5 had no home, so it lived on the link board as 145 dead rows | U5 |
-| 6 | **Benches** — registry, identity, enforced rule, migration | 69% link-board rot; the same bench posted 5× | U6 |
+| 6 | ~~**Benches**~~ — **landed `1c3ce5d`, released `v0.6.0`** | 69% link-board rot (re-measured: 178 booth rows + 8 bench re-posts) | U6 |
-| 7 | **Navigation at 270 items** — sections, rail, filters, grid keyboard | one flat wall; subfolder structure discarded at render | U7 |
+| 7 | ~~**Navigation**~~ — ~~sections~~ **filename groups**, rail, filters, grid keyboard — **landed, unreleased** | one flat wall; 0 of 11 galleries have subfolders, so grouping comes from the filename | U7 |
 
 Ordering is dependency-driven, not priority-driven: **U1 → U2 → {U3, U4, U5} →
 U7**, with **U6 independent** of all of them (different storage, different
 surface) and therefore the safest thing to land first or in parallel.
 
-**U1 and U2 are landed**, which unblocks U3, U4 and U5 — all three read marks.
+**ALL SEVEN UNITS ARE LANDED.** U7 closed last; its only dependency was
-**U5 is next** (operator, 2026-09-21). U6 remains independent and unstarted.
+`{U3, U4, U5}` and that closed with U3.
+
+⚠ **What U7 actually shipped is not what this row first described, and the
+difference is measured.** Sections were dropped for filename-prefix groups
+(operator-ratified 2026-09-22) because zero of eleven gallery booths have a
+subdirectory. Then the *grouping rule itself* changed at implementation: the
+contract's `strip a trailing digit run` yields 24 groups for `sindra-bakeoff`'s
+40 images and 27 for `sindra`'s 30 — a rail with a row per tile — because it
+keys on the end of the stem, where the instance number lives. The shipped rule
+keys on the **first separator-delimited segment**, where the family lives, and
+gives 4 and 2. The full re-measurement across all 17 live booths is in
+`docs/contracts/u7_navigation.contract.md`.
+
+**The v1 target is met.** What remains is a release decision the operator owns:
+cut `1.0`, or take a `0.7.0` staging release first. Nothing in the code is
+waiting on it.
+
+**U5's adoption is a measured prediction, not a finished result**, and it is
+TWO predictions rather than one. The operator declined a fleetwide announcement
+so that adoption could be told apart from design; within fifty minutes of the
+deploy a peer that had been told nothing (`comfy-dev`) created a booth and it
+announced itself with a handle and an empty `why`. That is the split:
+
+- **The handle rides for free.** It is written by `booth new` and `booth add`,
+  so every existing caller starts announcing without learning anything.
+- **The `why` has to be learned.** It needs someone to know the flag exists.
+
+Both get re-measured on **2026-09-29**:
+
+    find ~/booth-data -maxdepth 2 -name .booth.json | wc -l            # free
+    grep -l '"why": "[^"]' ~/booth-data/*/.booth.json | wc -l          # learned
+
+A high first count with a near-zero second is the predicted shape of "nobody was
+told" — an adoption failure fixed by announcing, which is a different thing from
+nobody wanting it. Same instrument as U4's `.forever` prediction below.
 
 ### Cross-cutting invariant — deterministic order, everywhere
 
@@ -53,11 +92,37 @@ Where it already binds, and what the rule is in each case:
 | marks in a booth | `(created, id)` — time, with the id as tie-break so two marks written in the same second cannot swap |
 | legacy ask import | `(mtime, name)`, which is the order `list_asks` gave them |
 | link board rows | pinned first, then newest-first |
+| a booth's announcement | not a collection — one flat record per booth, nothing to order (U5) |
+| **groups among themselves** | **the position of each group's first member in the rendered sequence** — `sorted(rel)` narrowed by the filter, never re-sorted. Walking the rendered list once into an insertion-ordered dict IS the rule, so there is no second sort to drift from it (U7) |
+| **items within a group** | not a separate order — a group is a label on a tile, not a container. The grid stays `sorted(rel)` and groups interleave in it freely (U7) |
+| the bench registry | `(state rank, name casefolded, id)` — live before promoted before retired, then alphabetical, with the id as a TOTAL tie-break so two benches sharing a name cannot swap (U6) |
+| the link board's dead marker | not an order — a per-row stamp read from the existing `order_for_display` sequence, so marking cannot move a row (U6) |
+| embed anchors in a verbatim report | **document order** — what `querySelectorAll` yields, so the author's markup decides (U3) |
+| the embed tail (fragments the author did not place) | **payload order**, which is the marks order `(created, id)` — one rule, whether a fragment lands at an anchor or at the end (U3) |
+| questions within a pick | declaration order, in the payload's `questions` LIST — carried by the format rather than by object-key insertion order (U3) |
 
-Where it is still to be decided, and must be before the unit ships: **U7's
+U3's three rows are the first case where the rule binds across a language
-section ordering and its compare pairing** (sections need a stated order among
+boundary: the order is decided in Python and honoured in JavaScript, and a
-themselves, not just within; pairing by filename needs a rule for what happens
+browser test asserts it rather than a string assertion that could not see it.
-to an unpaired file), and **U6's bench listing**.
+
+U4 added no ordered collection — a booth's lifetime is one state per booth,
+not a sequence — so the rule above did not need a new row. The three lifetime
+surfaces (index card, booth header, marks page) render through ONE macro
+precisely so they cannot disagree, which is the same property stated for
+ordering: one rule, one place, every surface reading it.
+
+**U7's group ordering is SETTLED and SHIPPED** (operator, 2026-09-22): groups
+order by the position of their first member in the rendered sequence, so the
+rail reads in the same direction as the grid. Subfolder sections were dropped
+in favour of filename-prefix groups on measured grounds — zero of eleven
+gallery booths have a subdirectory.
+
+**Nothing in this table is undecided any more.** The two U7 rules that were
+(section ordering among themselves, compare pairing) resolved differently:
+section ordering is MOOT, because U7 renders no section rail — `Item.section`
+still exists and is still derived, it simply has no ordered surface. Compare
+pairing rode into v1.1 with compare mode itself. **U6's bench listing is
+settled** — the row above.
 
 The test for any new ordered surface: *can you write the rule down in one line?*
 If not, it does not have one yet.

booth/__init__.py

@@ -1,3 +1,42 @@
-"""The Booth — ephemeral media drop board. See booth.app for the server."""
+"""The Booth — ephemeral media drop board. See booth.app for the server.
 
-__version__ = "0.1.0"
+⚠ THIS FILE IS EFFECTIVELY STDLIB-ONLY and nothing used to say so. `scripts/booth`
+imports `booth.links` / `booth.marks` / `booth.manifest` under the SYSTEM python3
+with no venv, and importing any of them executes this module first — so a single
+third-party import here breaks `booth ask` on every fleet host exactly as one in
+those three would. `test_stdlib_only` now covers `__init__` for that reason.
+"""
+
+import tomllib
+from pathlib import Path
+
+_PYPROJECT = Path(__file__).resolve().parent.parent / "pyproject.toml"
+
+
+def _declared_version() -> str:
+    """The version of the code actually running, read from `pyproject.toml`.
+
+    ⚠ NOT `importlib.metadata`, and the reason is this repo's own shape: there
+    is no build step and no install step — `booth.service` runs uvicorn with
+    WorkingDirectory set to the repo, so the running code IS this tree.
+    Installed metadata describes a DIFFERENT artifact and was found saying
+    `0.3.0` (a vestigial dist-info, three releases stale, with no package
+    directory behind it) while the tree was at `1.0.0b1`. A confidently wrong
+    number that varies by environment is worse than the hardcoded `0.1.0` this
+    replaced, which at least failed the same way everywhere.
+
+    Falls back to installed metadata for the case this repo does not have but a
+    consumer might: packaged as a wheel, where pyproject does not ship.
+    """
+    try:
+        return tomllib.loads(_PYPROJECT.read_text())["project"]["version"]
+    except (OSError, KeyError, tomllib.TOMLDecodeError):
+        try:
+            from importlib.metadata import version
+
+            return version("booth")
+        except Exception:
+            return "0.0.0+unknown"
+
+
+__version__ = _declared_version()

booth/benches.py

@@ -0,0 +1,416 @@
+"""Benches: a running thing, registered.
+
+A bench is NOT a booth and NOT a bookmark. It is a durable middle-to-long-term
+testing surface — jackdaw's current bench, talk's current bench, the things that
+get promoted to Homepage when they are fully deployed. The standing link board
+absorbed the job because it was the only surface on offer, and an O_APPEND log
+with no identity turns "here is the bench again" into a fifth row rather than an
+update: `talk` is on the board five times and Peedlar's root three.
+
+STDLIB ONLY, AND SIBLING-FREE, ON PURPOSE. `scripts/booth` imports this through
+a `python3 -c` heredoc under the system python3 with no venv, exactly as it
+imports `marks`, `asks`, `links` and `manifest`. A third-party import breaks
+`booth bench` on every fleet host; a `from booth.links import ...` breaks it on
+any host where both modules are not importable together, which is a second way
+for the same invariant to fall. `tests/test_benches.py` forbids both.
+
+SINGLE-WRITER, MANY-READER — the opposite shape from `links.md`. The board is a
+multi-writer append log because seventeen agent handles post to it at once. This
+is the operator in one browser plus occasional CLI calls, so it is one file,
+rewritten whole under a lock, replaced atomically. Inheriting the append-log
+design here would be the mistake CLAUDE.md names by name.
+"""
+
+from __future__ import annotations
+
+import fcntl
+import json
+import os
+import stat
+import tempfile
+from dataclasses import dataclass, replace
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Iterable
+from urllib.parse import urlsplit, urlunsplit
+
+# At the DATA ROOT, not inside a booth. A dotfile there is invisible to
+# `list_booths` and to `sweep_once` — both skip a child that is not a directory
+# AND a child whose name starts with a dot, so the registry fails two guards
+# rather than one. Verified against both functions (seam review SR-4, SR-5)
+# rather than assumed: had either guard been absent, the sweeper would have
+# eaten this file on its first tick.
+BENCHES_FILE = ".benches.json"
+BENCH_LOCK = ".benches.lock"
+
+# live → promoted (to Homepage) → retired. Order is meaningful: it is the
+# first key of the rendered order, so a retired bench sinks.
+BENCH_STATES = ("live", "promoted", "retired")
+_STATE_RANK = {s: i for i, s in enumerate(BENCH_STATES)}
+
+# Display budgets, not storage limits — these land in a panel row.
+NAME_MAX, OWNER_MAX, URL_MAX = 120, 64, 2048
+
+# The read is on the render path, so it is bounded. 256 KiB holds thousands of
+# benches; the live board has 43 non-booth rows total.
+BENCHES_MAX_BYTES = 256 * 1024
+
+_SCHEMES = ("http", "https")
+
+
+@dataclass(frozen=True)
+class Bench:
+    """One registered bench.
+
+    `id` and `url` are two fields ON PURPOSE. The identity must be normalized so
+    that re-posting updates rather than appends; the href must be verbatim so a
+    server that cares about a trailing slash, a case-sensitive path or a query
+    still works when the operator clicks it. Collapsing them would make the
+    registry quietly change where a link goes — a bug that surfaces as "the
+    bench 404s" and is never traced back here.
+    """
+
+    id: str                    # the normalized URL — identity, and the key on disk
+    url: str                   # the URL as posted — what a click goes to
+    name: str
+    owner: str                 # an althing handle, or "booth" for the service
+    state: str
+    added: str                 # ISO-8601 with offset, from the FIRST registration
+    updated: str               # ISO-8601 with offset, from the most recent upsert
+    error: str | None = None   # a read-time verdict; never stored
+
+
+def normalize_bench_url(url: str) -> str:
+    """The identity of a bench. Raises ValueError with a reason a human can act on.
+
+    THE RULE, in full, because a vague identity is worse than a wrong one:
+
+      * surrounding whitespace stripped
+      * scheme lowercased; anything but http/https refused
+      * userinfo (`user:pass@host`) REFUSED, never stripped
+      * host lowercased; an empty host refused
+      * port dropped when it is the scheme default (80 http, 443 https)
+      * path kept verbatim, except that a bare "/" becomes ""
+      * query kept verbatim INCLUDING parameter order (a query is opaque)
+      * fragment dropped
+
+    WHY THE FULL URL AND NOT THE ORIGIN — measured, not chosen. Collapsing the
+    live board's 43 non-booth rows by origin yields 19 groups; by full URL, 35.
+    The difference is not duplication: it is eight distinct gitea repositories
+    merged into one row, three unrelated HuggingFace model cards merged into
+    one, and the two LRPG surfaces on `10.100.10.50:8321` merged into one —
+    which are the information-architecture doc's own example of two real
+    benches. Origin identity destroys more than it deduplicates. Full-URL
+    identity still collapses both cases that doc names: talk 5 → 1, Peedlar 3 → 1.
+
+    WHY THE QUERY IS IN AND THE FRAGMENT IS OUT. Three ShutterChute rows on the
+    board differ only by `?token=`; they are three genuinely different one-shot
+    links, and dropping the query would merge them into a bench that is none of
+    them. A fragment is a position inside a page, never a different resource.
+    """
+    raw = (url or "").strip()
+    if not raw:
+        raise ValueError("a bench needs a URL")
+    if len(raw) > URL_MAX:
+        raise ValueError(f"URL is longer than {URL_MAX} characters")
+    try:
+        parts = urlsplit(raw)
+    except ValueError as exc:                       # malformed IPv6 literal, etc.
+        raise ValueError(f"could not parse that URL: {exc}") from exc
+
+    scheme = parts.scheme.lower()
+    if scheme not in _SCHEMES:
+        raise ValueError(
+            f"a bench must be http or https, not {parts.scheme or '(no scheme)'}"
+        )
+    if "@" in parts.netloc:
+        # Refused, NOT stripped. Stripping would register a bench whose URL no
+        # longer works while telling the poster it succeeded — and would put a
+        # credential on a board that renders on an unauthenticated LAN surface
+        # on the way there.
+        raise ValueError("a bench URL must not carry credentials; strip the user:pass@ and re-post")
+    try:
+        host = (parts.hostname or "").lower()
+        port = parts.port
+    except ValueError as exc:                       # a non-numeric port
+        raise ValueError(f"could not read the host or port: {exc}") from exc
+    if not host:
+        raise ValueError("that URL has no host")
+
+    # RE-WRAP A BRACKETED IPv6 LITERAL. `urlsplit().hostname` strips the
+    # brackets, and rebuilding the netloc from it produces `http://::1:8080/a`
+    # — not a different spelling of the same URL but a BROKEN one, so a re-post
+    # never matches the row the operator thinks they are updating. The bracket
+    # is part of the authority's syntax, not decoration. Detected by the colon,
+    # which cannot appear in a hostname or an IPv4 literal.
+    if ":" in host:
+        host = f"[{host}]"
+    default = {"http": 80, "https": 443}[scheme]
+    netloc = host if port in (None, default) else f"{host}:{port}"
+    # A bare "/" is the same resource as no path at all; a trailing slash on a
+    # REAL path is not, and is left alone.
+    path = "" if parts.path == "/" else parts.path
+    return urlunsplit((scheme, netloc, path, parts.query, ""))
+
+
+# ---- storage ----------------------------------------------------------------
+
+
+def _now() -> str:
+    return datetime.now(timezone.utc).isoformat(timespec="seconds")
+
+
+def _cap(value: object, limit: int, field: str) -> str:
+    if not isinstance(value, str):
+        raise ValueError(f"{field} must be text, not {type(value).__name__}")
+    return value[:limit]
+
+
+def _bench_from(bench_id: str, row: object) -> Bench:
+    """One stored row to a record. Raises ValueError on any shape it cannot
+    trust — this is the STRICT half, used by the write path and by the read
+    path's single try/except."""
+    if not isinstance(row, dict):
+        raise ValueError(f"{bench_id}: expected an object, found {type(row).__name__}")
+    state = row.get("state", "live")
+    if state not in BENCH_STATES:
+        raise ValueError(f"{bench_id}: unknown state {state!r}")
+    url = row.get("url", bench_id)
+    if not isinstance(url, str):
+        raise ValueError(f"{bench_id}: url must be text, not {type(url).__name__}")
+    if len(url) > URL_MAX:
+        # REFUSED, NOT TRUNCATED — unlike `name` and `owner`. Those are display
+        # budgets and clipping one costs a few characters in a panel row. A
+        # clipped URL is a DEAD ANCHOR, and INV-7 promises the click goes to the
+        # posted address byte for byte; silently shortening it keeps the promise
+        # in the type system and breaks it in the browser. Nothing this code
+        # writes can get here (normalize refuses over-long input); a hand-edited
+        # registry can, and it is damage, which is what the reader reports.
+        raise ValueError(f"{bench_id}: url is longer than {URL_MAX} characters")
+    return Bench(
+        id=bench_id,
+        url=url,
+        name=_cap(row.get("name", ""), NAME_MAX, "name"),
+        owner=_cap(row.get("owner", ""), OWNER_MAX, "owner"),
+        state=state,
+        added=_cap(row.get("added", ""), 64, "added"),
+        updated=_cap(row.get("updated", ""), 64, "updated"),
+    )
+
+
+def _read_bytes(path: Path) -> bytes:
+    """Read at most BENCHES_MAX_BYTES + 1 bytes from a REGULAR FILE.
+
+    REGULAR-FILE FIRST, THEN SIZE, THEN A BOUNDED READ — in that order, and the
+    order is the whole point. A named pipe blocks in `open()`, before any byte
+    cap can apply: bounding the read does NOT close that hole, and an earlier
+    draft of this module claimed it did while hanging on the first FIFO put at
+    this path. `read_benches` is on the board page's render path, so that hang
+    is a request that never returns and, with enough of them, the threadpool
+    behind every route. `marks.py` learned this on 2026-09-22 and guards with
+    `S_ISREG`; this is the same guard, not a new idea.
+
+    The bounded read stays, for the case the stat cannot answer: a regular file
+    that GREW between the stat and the read.
+    """
+    st = os.stat(path)
+    if not stat.S_ISREG(st.st_mode):
+        raise ValueError(f"{path.name} is not a regular file")
+    if st.st_size > BENCHES_MAX_BYTES:
+        raise ValueError(f"registry is larger than {BENCHES_MAX_BYTES} bytes")
+    with path.open("rb") as fh:
+        return fh.read(BENCHES_MAX_BYTES + 1)
+
+
+def _load_strict(root: Path) -> dict[str, Bench]:
+    """Every bench, or ValueError. The write path's reader.
+
+    Whole-file, not per-row: a registry with one unreadable row is a registry
+    somebody has to look at, and quietly dropping the row is how a bench
+    disappears without anyone being told.
+    """
+    path = Path(root) / BENCHES_FILE
+    if not path.exists():
+        return {}
+    blob = _read_bytes(path)
+    if len(blob) > BENCHES_MAX_BYTES:
+        raise ValueError(f"registry is larger than {BENCHES_MAX_BYTES} bytes")
+    try:
+        raw = json.loads(blob.decode("utf-8"))
+    except (UnicodeDecodeError, json.JSONDecodeError) as exc:
+        raise ValueError(f"registry is not valid JSON: {exc}") from exc
+    if not isinstance(raw, dict):
+        raise ValueError(f"registry must be an object keyed by URL, found {type(raw).__name__}")
+    return {k: _bench_from(k, v) for k, v in raw.items()}
+
+
+def read_benches(root: Path) -> tuple[list[Bench], str | None]:
+    """Every registered bench in the rendered order, plus a read-time error.
+
+    NEVER RAISES. This runs on the render path, and the v0.2.2 lesson in this
+    repo was learned the expensive way: a poisoned `.marks.json` returned 500
+    for `/` and `/healthz` across all 25 booths. A registry that cannot be read
+    costs its own panel, never the page.
+
+    ABSENT AND DAMAGED ARE DIFFERENT and must render differently — only one of
+    them needs a human. Absent is `([], None)`; damaged is `([], "why")`.
+    """
+    try:
+        return order_benches(_load_strict(root).values()), None
+    except ValueError as exc:
+        return [], str(exc)
+    except OSError as exc:
+        return [], f"registry could not be read: {exc}"
+    except RecursionError:
+        # Deeply nested JSON (`[[[[...`) blows the stack inside json.loads, and
+        # RecursionError is neither ValueError nor OSError — so it escaped the
+        # pair above and 500'd the page this function exists to protect. The
+        # byte cap does not help: 200k open brackets is 200 KB.
+        return [], "registry is nested too deeply to parse"
+
+
+def _write_all(root: Path, benches: dict[str, Bench]) -> None:
+    """Atomic replace. Caller holds the lock.
+
+    Temp file + os.replace, so a reader never sees a partial file and a crash
+    mid-write cannot truncate the registry into a shorter — and therefore
+    quieter — set of benches. CLAUDE.md invariant 5.
+    """
+    root = Path(root)
+    path = root / BENCHES_FILE
+    payload = {
+        b.id: {"url": b.url, "name": b.name, "owner": b.owner,
+               "state": b.state, "added": b.added, "updated": b.updated}
+        # The key IS the id, so the record does not carry it twice — two copies
+        # of one fact is two things that can disagree.
+        for b in benches.values()
+    }
+    # Per-pid scratch name so two writers cannot share it: the atomic-replace
+    # promise is that a READER never sees a partial file, not that two writers
+    # never collide on the way there.
+    body = json.dumps(payload, indent=2, sort_keys=True) + "\n"
+    # THE WRITER RESPECTS THE READER'S CAP. Without this, a successful
+    # registration can push the file past BENCHES_MAX_BYTES and every
+    # subsequent read fails — so the LAST bench somebody added is the one that
+    # makes all the others invisible, and the write that did it reported
+    # success. The reader is lenient about damage; it is not lenient about
+    # size, and a writer that ignores a limit its own reader enforces is
+    # manufacturing exactly the state the leniency exists to survive.
+    if len(body.encode("utf-8")) > BENCHES_MAX_BYTES:
+        raise ValueError(
+            f"that registration would push the registry past {BENCHES_MAX_BYTES} "
+            f"bytes, which its own reader refuses; nothing was written")
+    # AN UNPREDICTABLE SCRATCH NAME, IN THE SAME DIRECTORY. `.tmp.<pid>` is
+    # guessable, and a pre-planted symlink there redirects the write straight
+    # through the atomic replace — the replace is atomic, not safe. mkstemp
+    # creates with O_EXCL and 0600, so it cannot land on someone else's file.
+    # Same directory because os.replace is only atomic within a filesystem.
+    fd, tmpname = tempfile.mkstemp(dir=str(root), prefix=".benches-", suffix=".tmp")
+    tmp = Path(tmpname)
+    try:
+        with os.fdopen(fd, "w", encoding="utf-8") as fh:
+            fh.write(body)
+            fh.flush()
+            # FSYNC BEFORE THE REPLACE. os.replace orders the rename, not the
+            # DATA behind it: without this, a power loss can publish a name
+            # pointing at bytes that never reached the disk, which is a
+            # truncated registry wearing a successful write's clothes.
+            os.fsync(fh.fileno())
+        os.chmod(tmp, 0o644)          # mkstemp's 0600 is tighter than the rest
+        os.replace(tmp, path)
+    except BaseException:
+        # A write that dies between create and replace would otherwise strand
+        # the scratch file beside the registry forever. The prior registry is
+        # untouched either way — os.replace is the only thing that publishes.
+        tmp.unlink(missing_ok=True)
+        raise
+
+
+class _Locked:
+    """Exclusive flock over the whole read-modify-write, on a sidecar."""
+
+    def __init__(self, root: Path):
+        self.root = Path(root)
+        self.root.mkdir(parents=True, exist_ok=True)
+        self.path = self.root / BENCH_LOCK
+
+    def __enter__(self):
+        self.path.touch(exist_ok=True)
+        self.fh = self.path.open("r+")
+        fcntl.flock(self.fh, fcntl.LOCK_EX)
+        return self
+
+    def __exit__(self, *exc):
+        fcntl.flock(self.fh, fcntl.LOCK_UN)
+        self.fh.close()
+        return False
+
+
+def upsert_bench(root: Path, url: str, name: str, owner: str) -> tuple[Bench, bool]:
+    """Register or update by normalized URL. Returns (bench, created).
+
+    READS ARE LENIENT, WRITES ARE STRICT — and this is the strict side. A write
+    over a registry that cannot be parsed RAISES rather than starting a fresh
+    one: on 2026-09-21 this repo learned that a tolerant writer over a damaged
+    `.marks.json` wipes the operator's judgment, and a tolerant reader is a
+    completely different decision from a tolerant writer.
+
+    `added` survives an update; `state` survives too, so a promoted bench that
+    re-announces itself after a deploy is not silently demoted.
+    """
+    bench_id = normalize_bench_url(url)
+    with _Locked(root):
+        benches = _load_strict(root)          # raises on damaged — deliberate
+        prior = benches.get(bench_id)
+        now = _now()
+        bench = Bench(
+            id=bench_id,
+            url=(url or "").strip(),
+            name=_cap(name or "", NAME_MAX, "name"),
+            owner=_cap(owner or "", OWNER_MAX, "owner"),
+            state=prior.state if prior else "live",
+            added=prior.added if prior else now,
+            updated=now,
+        )
+        benches[bench_id] = bench
+        _write_all(root, benches)
+    return bench, prior is None
+
+
+def set_bench_state(root: Path, bench_id: str, state: str) -> Bench | None:
+    """Move a bench between live / promoted / retired. None if no such bench."""
+    if state not in BENCH_STATES:
+        raise ValueError(f"state must be one of {', '.join(BENCH_STATES)}, not {state!r}")
+    with _Locked(root):
+        benches = _load_strict(root)
+        prior = benches.get(bench_id)
+        if prior is None:
+            return None
+        moved = replace(prior, state=state, updated=_now())
+        benches[bench_id] = moved
+        _write_all(root, benches)
+    return moved
+
+
+def remove_bench(root: Path, bench_id: str) -> Bench | None:
+    """Drop one bench. Returns the removed record, or None."""
+    with _Locked(root):
+        benches = _load_strict(root)
+        gone = benches.pop(bench_id, None)
+        if gone is None:
+            return None
+        _write_all(root, benches)
+    return gone
+
+
+def order_benches(benches: Iterable[Bench]) -> list[Bench]:
+    """ORDER: (state rank, name casefolded, id).
+
+    live before promoted before retired, then alphabetical, with the id as a
+    TOTAL tie-break so two benches sharing a name cannot swap between renders.
+    CLAUDE.md invariant 6 — the Booth's job is comparison, and an order that
+    moves between page loads files the operator's judgment against the wrong
+    row. Pure: no I/O, and the input sequence is not mutated.
+    """
+    return sorted(benches, key=lambda b: (_STATE_RANK.get(b.state, len(BENCH_STATES)),
+                                          b.name.casefold(), b.id))

booth/inline.py

@@ -1,119 +0,0 @@
-"""Inline ask placement inside a booth's VERBATIM index.html.
-
-A booth that ships its own `index.html` is served untouched, so the auto-gallery
-template's asks panel never renders there. The first fix was a chip linking to a
-separate `/asks` page; the operator's verdict on that (2026-09-09) was that the
-question belongs WITH the artifacts it is about — a four-voice audition wants the
-radio group for each voice under that voice's audio, not on another page.
-
-So the report author marks where each piece goes, with a placeholder element:
-
-    <div data-booth-ask="anchors"></div>          the whole ask: every question + submit
-    <div data-booth-ask="anchors:lawson"></div>   just that question's radios
-    <div data-booth-ask-submit="anchors"></div>   the notes field + submit button
-
-Per-question fragments bind to ONE form via the HTML5 `form=` attribute, so four
-groups scattered down a page still submit as a single POST — which is what a
-multi-question ask requires (every question or 400). No JavaScript.
-
-An `<!-- booth:ask anchors -->` comment works the same way, for authors who would
-rather not put an empty div in their markup.
-
-Placement is OPTIONAL. A page with no placeholders gets the whole ask appended at
-the end of its body, so an ask is never invisible — that guarantee is the point,
-and marking it up only moves it somewhere better.
-"""
-
-from __future__ import annotations
-
-import re
-
-# <div data-booth-ask="stem"></div>  /  <span data-booth-ask="stem:key"></span>
-_EL_RE = re.compile(
-    r"<(?P<tag>[A-Za-z][\w-]*)\b[^>]*?\bdata-booth-ask=\"(?P<spec>[^\"]+)\"[^>]*?>"
-    r"(?:\s*</(?P=tag)\s*>)?",
-    re.IGNORECASE,
-)
-_SUBMIT_EL_RE = re.compile(
-    r"<(?P<tag>[A-Za-z][\w-]*)\b[^>]*?\bdata-booth-ask-submit=\"(?P<spec>[^\"]+)\"[^>]*?>"
-    r"(?:\s*</(?P=tag)\s*>)?",
-    re.IGNORECASE,
-)
-# <!-- booth:ask stem --> / <!-- booth:ask stem:key --> / <!-- booth:ask-submit stem -->
-_COMMENT_RE = re.compile(r"<!--\s*booth:ask\s+(?P<spec>[^\s>-][^\s>]*)\s*-->", re.IGNORECASE)
-_COMMENT_SUBMIT_RE = re.compile(r"<!--\s*booth:ask-submit\s+(?P<spec>[^\s>]+)\s*-->", re.IGNORECASE)
-
-
-def split_spec(spec: str) -> tuple[str, str | None]:
-    """`"anchors:lawson"` -> `("anchors", "lawson")`; `"anchors"` -> `("anchors", None)`."""
-    stem, sep, key = spec.strip().partition(":")
-    return stem.strip(), (key.strip() or None) if sep else None
-
-
-def has_placeholders(html: str) -> bool:
-    return bool(
-        _EL_RE.search(html) or _SUBMIT_EL_RE.search(html)
-        or _COMMENT_RE.search(html) or _COMMENT_SUBMIT_RE.search(html)
-    )
-
-
-def form_id(stem: str) -> str:
-    return f"bk-ask-form-{re.sub(r'[^A-Za-z0-9_-]', '-', stem)}"
-
-
-def place(html: str, asks: list, render) -> tuple[str, dict[str, set], set[str]]:
-    """Substitute every placeholder with rendered ask HTML.
-
-    `render(kind, ask, key)` returns the fragment for kind in
-    {"whole", "question", "submit"}. Returns the new html; a map of stem ->
-    the set of question keys placed inline (with `None` in the set meaning the
-    WHOLE ask was placed); and the set of stems whose submit block was placed
-    explicitly.
-
-    The caller needs the per-key detail, not just "this stem appeared
-    somewhere": a multi-question ask requires EVERY question on submit, so a
-    page that marks up two of four questions must still be handed the other two
-    or the form is unsubmittable — a 400 the operator would meet only after
-    filling it in.
-
-    A placeholder naming an ask this booth does not have is left ALONE, not
-    blanked: silently eating the author's markup would hide a typo'd stem, and
-    an untouched empty div is invisible anyway.
-    """
-    # Marks index by ATTRIBUTE, not subscript: `place` was the one consumer in
-    # the service that did `a["stem"]`, which a frozen dataclass refuses. Caught
-    # by the U2 seam review (SR-1) — the cold contract pass cannot see a sibling
-    # module's surface by design, so nothing else would have found it before the
-    # first verbatim booth 500'd.
-    by_stem = {a.id: a for a in asks}
-    placed: dict[str, set] = {}
-    submitted: set[str] = set()
-
-    def sub_main(m: re.Match) -> str:
-        stem, key = split_spec(m.group("spec"))
-        ask = by_stem.get(stem)
-        if ask is None:
-            return m.group(0)
-        if key is None:
-            placed.setdefault(stem, set()).add(None)
-            submitted.add(stem)
-            return render("whole", ask, None)
-        q = next((q for q in ask.questions if q.get("key") == key), None)
-        if q is None:
-            return m.group(0)
-        placed.setdefault(stem, set()).add(key)
-        return render("question", ask, key)
-
-    def sub_submit(m: re.Match) -> str:
-        stem, _ = split_spec(m.group("spec"))
-        ask = by_stem.get(stem)
-        if ask is None:
-            return m.group(0)
-        placed.setdefault(stem, set())
-        submitted.add(stem)
-        return render("submit", ask, None)
-
-    for pat, fn in ((_EL_RE, sub_main), (_COMMENT_RE, sub_main),
-                    (_SUBMIT_EL_RE, sub_submit), (_COMMENT_SUBMIT_RE, sub_submit)):
-        html = pat.sub(fn, html)
-    return html, placed, submitted

booth/items.py

@@ -15,6 +15,7 @@ See docs/contracts/u1_item_record.contract.md.
 
 from __future__ import annotations
 
+import re
 from dataclasses import dataclass
 from pathlib import Path
 from typing import Sequence
@@ -89,6 +90,7 @@ class Item:
     url: str
     kind: str
     section: str | None
+    group: str | None
     caption: str | None
     blurred: bool
     doc: str | None
@@ -115,6 +117,43 @@ def _section_of(rel: str) -> str | None:
     return None if str(parent) == "." else parent.as_posix()
 
 
+# One separator run between name segments. A filename is the only grouping
+# signal the live booths actually carry: 0 of 11 galleries have a subdirectory.
+_SEG = re.compile(r"[-_. ]+")
+
+
+def _group_of(rel: str) -> str | None:
+    """The grouping key for an item, or None when it has none.
+
+    THE RULE, in one line: **the first separator-delimited segment of the
+    basename's stem — with a trailing digit run stripped only when the stem has
+    no separator at all.** `00-sheet-c1-market-noon.png` -> `00`;
+    `m-c1-market-noon-9401.png` -> `m`; `flag-rear.png` -> `flag`;
+    `ac01.png` -> `ac` (no separator, so the digits are the separator);
+    `v30-seed8302.png` -> `v30` (separator present, so `v30` survives and does
+    not merge with `v35`, which is the axis that booth is about).
+
+    None for a stem with nothing before the digits -- `01.png` has no prefix to
+    group on, and inventing one would file every numbered render under the
+    empty string.
+
+    ⚠ THIS IS NOT THE RULE THE CONTRACT FIRST NAMED. `strip ONE trailing run of
+    digits` was measured against the live set on 2026-09-22 and yields 24 groups
+    for sindra-bakeoff's 40 images and 27 for sindra's 30 -- a rail with one row
+    per tile. The contract's own table claimed 5 and 1 for those two booths;
+    neither reproduces under the rule it states beside them. The rewritten table
+    carries the re-measurement.
+
+    Derived HERE and nowhere else (INV-1). A route body that re-derived it would
+    be the caption bug in a new field.
+    """
+    stem = Path(rel).stem  # basename without its last suffix; `a.tar.gz` -> `a.tar`
+    segs = _SEG.split(stem)
+    if len(segs) == 1:
+        return re.sub(r"\d+$", "", stem) or None
+    return segs[0] or None
+
+
 def _resolve_captions(by_rel: dict[str, Path]) -> tuple[dict[str, str], set[str]]:
     """(caption-by-rel, rels consumed as sidecars).
 
@@ -203,6 +242,7 @@ def booth_items(booth: Path) -> list[Item]:
                 url=quote(rel, safe="/"),
                 kind=classify(p.name),
                 section=_section_of(rel),
+                group=_group_of(rel),
                 caption=caption.get(rel),
                 blurred=rel in blurred,
                 doc=doc_kind(p.name),

booth/links.py

@@ -14,6 +14,7 @@ import hashlib
 import os
 import re
 from pathlib import Path
+from urllib.parse import unquote, urlsplit
 
 # ---- the standing link board ------------------------------------------------
 #
@@ -194,3 +195,58 @@ def order_for_display(entries: list[dict], pinned: set[str]) -> list[dict]:
     stamped = [{**e, "pinned": e["id"] in pinned} for e in entries]
     stamped.reverse()  # newest first
     return [e for e in stamped if e["pinned"]] + [e for e in stamped if not e["pinned"]]
+
+
+# ---- what counts as a booth link -------------------------------------------
+
+
+def booth_target(url: str) -> str | None:
+    """The booth NAME a URL points at, or None when it is not a booth link.
+
+    ONE PREDICATE, THREE CALLERS — the CLI's `link` refusal, the board's
+    dead-row marker, and `bench import`'s classifier. They must agree: a rule
+    that refuses a shape the board then fails to mark as dead (or the reverse)
+    is two readers of one truth, which is the bug this repo has now paid for
+    three times. `tests/test_benches.py` runs one table through every caller.
+
+    HOST-AGNOSTIC AND PATH-SHAPED. A row is a booth link when its path is
+    `/b/<name>` or `/b/<name>/...`, whatever the host. NOT a host allowlist: the
+    fleet reaches this service as `10.100.10.50:8090`, `localhost:8090` and
+    `nh3-dev.nh3.internal:8090`, and an allowlist would silently fail to refuse
+    from whichever name somebody used next — a rule that fails OPEN on the exact
+    case it exists to catch. The accepted cost is that a third-party URL with a
+    `/b/<x>` path reads as a booth link; that failure is visible (a refusal
+    naming the reason) rather than silent, and no such URL is on the board.
+
+    THE NAME SEGMENT IS PERCENT-DECODED. `app.py` emits booth links through
+    `quote(name, safe="")`, so a booth whose name needs encoding appears on the
+    board encoded. Comparing the raw segment against a directory name would mark
+    every such booth permanently dead and echo the encoded form back at the
+    poster in the refusal message.
+
+    The returned name passes the SAME addressability rules `resolve_booth`
+    enforces (non-empty, no leading dot, no separator, no `..`), so the two
+    cannot disagree about what is reachable.
+
+    NEVER RAISES. A board row is arbitrary operator-editable text; a predicate
+    that raises on one row takes the whole page.
+    """
+    try:
+        parts = urlsplit((url or "").strip())
+        if parts.scheme.lower() not in ("http", "https"):
+            return None
+        segments = parts.path.split("/")
+        if len(segments) < 3 or segments[1] != "b":
+            return None
+        name = unquote(segments[2])
+    except (ValueError, UnicodeDecodeError):
+        return None
+    if not name or name.startswith(".") or "/" in name or "\\" in name or ".." in name:
+        return None
+    # `unquote` will happily hand back a NUL or a newline, and neither can name
+    # a directory. Unfiltered they reach `is_dir()` (ValueError on an embedded
+    # NUL, which is NOT an OSError and so escapes the marker's guard), the
+    # refusal message the CLI prints, and the marker the board renders.
+    if any(ch in name for ch in "\x00") or any(ord(ch) < 0x20 for ch in name):
+        return None
+    return name

booth/manifest.py

@@ -0,0 +1,274 @@
+"""A booth's own announcement — who posted it, and why.
+
+U5. The index card used to show a name, an item count and a countdown, and
+nothing the poster chose. An agent with something to show therefore had no way
+to make the booth say "look at this" and posted a URL to the link board
+instead — which is why 145 of that board's 210 rows (69%) ended up pointing at
+booths that had already been swept. The board was absorbing a job it was never
+shaped for. This is the shape.
+
+    .booth.json  ->  {"handle": ..., "title": ..., "why": ..., "created": ...}
+
+⚠ STDLIB ONLY, and it imports nothing from `booth.*` either.
+
+`scripts/booth` — the CLI every fleet session uses — imports this module
+directly under the system `python3` with no venv, through a `python3 -c`
+heredoc no AST extractor can see. A single third-party import here breaks
+`booth new` and `booth add` on every host, and the failure surfaces in an
+agent's session rather than in ours. The ban extends to sibling `booth` modules:
+importing `marks` to reuse its atomic write would drag marks' own import list
+into this one's, so the four-line pattern is copied instead. `test_stdlib_only`
+in tests/test_manifest.py is the only thing standing here.
+
+Contract: docs/contracts/u5_booth_manifest.contract.md.
+"""
+from __future__ import annotations
+
+import json
+import os
+import secrets
+import stat as statmod
+from dataclasses import dataclass
+from datetime import datetime
+from pathlib import Path
+
+MANIFEST_FILE = ".booth.json"
+
+# A `why` renders inside a card's sub-line, so it is one line by construction
+# rather than by convention — enforced at the WRITE so nothing downstream has to
+# remember. The caps are display budgets, not storage limits.
+HANDLE_MAX = 64
+TITLE_MAX = 120
+WHY_MAX = 200
+CREATED_MAX = 64
+
+# A manifest is four short fields. Anything near this is not one, and reading it
+# into memory to find that out is the wrong order of operations: `list_booths`
+# calls the reader once per booth on every index load, so an unbounded read is
+# the service-wide outage the lenient reader exists to prevent, arriving in a
+# different costume. Checked by `stat`, before the bytes are touched.
+MANIFEST_MAX_BYTES = 64 * 1024
+
+# Where bytes that could not be read go when a re-announcement replaces them.
+# ONE fixed name, deliberately: a timestamped quarantine accumulates forever in
+# a folder nothing prunes, and the most recent damage is the only copy anybody
+# would look at. A dotfile, so it is invisible to every listing and zip.
+QUARANTINE_FILE = ".booth.json.broken"
+
+# The handle a booth created by the service itself carries. A pickup booth and
+# the standing link board are made by the Booth, not by an agent, and saying so
+# is true rather than manufactured — which is the whole reason there is no
+# exemption list. One rule: a booth with no manifest is unannounced.
+SERVICE_HANDLE = "booth"
+
+
+@dataclass(frozen=True)
+class Manifest:
+    """One booth's announcement.
+
+    `handle` is an althing agent handle, or `SERVICE_HANDLE` for a booth the
+    Booth made. `error` is a read-time verdict and is never stored.
+    """
+
+    handle: str
+    title: str
+    why: str
+    created: str
+    error: str | None = None
+
+
+def _one_line(value, limit: int) -> str:
+    """One line, bounded. Collapses ALL runs of whitespace, not only newlines —
+    a tab or a forty-space indent in a `why` renders as badly inside a card's
+    sub-line as a newline does, and the field is one line by construction."""
+    if not isinstance(value, str):
+        return ""
+    return " ".join(value.split())[:limit]
+
+
+def _temp_path(booth: Path) -> Path:
+    """A scratch name no other writer will pick.
+
+    Every writer used to derive the same `.booth.json.tmp`, so two `booth add`
+    calls on one booth could interleave through a stale descriptor into the
+    published path. Marks are protected from that by their flock; the manifest
+    deliberately has none — it is written once at creation, not read-modify-
+    written per click — so uniqueness is what stands in for the lock. Still a
+    dotfile, so no listing, gallery or zip can see it mid-write.
+    """
+    return booth / f"{MANIFEST_FILE}.{secrets.token_hex(4)}.tmp"
+
+
+def _as_doc(m: "Manifest") -> dict:
+    """The stored shape of a record, for the no-op comparison."""
+    return {"handle": m.handle, "title": m.title, "why": m.why, "created": m.created}
+
+
+def _now() -> str:
+    return datetime.now().astimezone().isoformat(timespec="seconds")
+
+
+def read_manifest(booth: Path) -> Manifest | None:
+    """This booth's announcement, or None if it never made one.
+
+    LENIENT, AND IT NEVER RAISES (INV-2). `list_booths` calls this once per
+    booth on every index page load, so a read that can raise is a service-wide
+    outage wearing a single-booth bug's clothes. That is not hypothetical: a
+    poisoned `.marks.json` did exactly that to `/` and `/healthz` across all 25
+    live booths, and the fix shipped in v0.2.2. Same posture, applied before the
+    same mistake rather than after it.
+
+    Absent -> None. Present but unreadable -> a Manifest carrying `error`, so a
+    card can say `unreadable` instead of quietly showing the same thing as a
+    booth that never announced (INV-5). Folding the two together would hide the
+    one case somebody has to go and fix.
+
+    Only `handle` is required. A hand-written manifest is a supported input —
+    the file is plain JSON in a folder the operator owns, and half the point of
+    the Booth is that a booth is just a directory.
+    """
+    booth = Path(booth)
+    path = booth / MANIFEST_FILE
+    # BOUNDED BEFORE THE READ. "Never raises" was not true of an unbounded one:
+    # a 4 GB file raises MemoryError and a deeply nested document raises
+    # RecursionError out of `json.loads`, and neither is an OSError or a
+    # ValueError. Both escape into `list_booths`, which calls this per booth on
+    # every index load — so one file returns 500 for the whole front page. Size
+    # first, by `stat`; then catch the two classes anyway, because a bound that
+    # is one day raised should not quietly re-open the hole.
+    try:
+        st = path.stat()
+    except FileNotFoundError:
+        return None
+    except OSError as exc:
+        return _broken(booth, f"cannot be read: {exc}")
+    # ⚠ REGULAR-FILE FIRST, then size. `st_size` answers a different question
+    # than "can this be read": it is 0 for a FIFO and 0 for /dev/zero, so both
+    # sail under the cap, and then `read_text` either blocks forever with no EOF
+    # or allocates until the kernel intervenes. The bound ABOVE is what made
+    # this reachable — a cap that trusts st_size inherits everything st_size
+    # does not mean. One such file stalls every `GET /` and `/healthz`.
+    if not statmod.S_ISREG(st.st_mode):
+        return _broken(booth, "is not a regular file")
+    if st.st_size > MANIFEST_MAX_BYTES:
+        return _broken(booth, f"is too large to be a manifest ({st.st_size} bytes)")
+    try:
+        text = path.read_text(encoding="utf-8")
+    except FileNotFoundError:
+        return None
+    except (OSError, UnicodeDecodeError, MemoryError) as exc:
+        return _broken(booth, f"cannot be read: {exc}")
+    if not text.strip():
+        return _broken(booth, "is empty")
+    try:
+        raw = json.loads(text)
+    except (ValueError, RecursionError, MemoryError) as exc:
+        return _broken(booth, f"is not valid JSON: {type(exc).__name__}")
+    if not isinstance(raw, dict):
+        return _broken(booth, "is not a JSON object")
+
+    handle = _one_line(raw.get("handle"), HANDLE_MAX)
+    if not handle:
+        return _broken(booth, "names no handle")
+    return Manifest(
+        handle=handle,
+        # `or booth.name` goes THROUGH the normalizer too. A directory name may
+        # legally carry a newline on POSIX and may run to 255 bytes, and the
+        # fallback used to hand either straight into a card's sub-line.
+        title=_one_line(raw.get("title"), TITLE_MAX) or _one_line(booth.name, TITLE_MAX),
+        why=_one_line(raw.get("why"), WHY_MAX),
+        created=_one_line(raw.get("created"), CREATED_MAX),
+    )
+
+
+def _broken(booth: Path, reason: str) -> Manifest:
+    # The directory name goes through the normalizer here too. This was the
+    # THIRD fallback of three; the write path's and the read path's were fixed a
+    # round earlier and this one was missed, with the same consequence — a
+    # newline or 255 bytes of directory name straight into a card's sub-line.
+    return Manifest(handle="", title=_one_line(booth.name, TITLE_MAX), why="",
+                    created="", error=f"{MANIFEST_FILE} {reason}")
+
+
+def write_manifest(booth: Path, handle: str, *, title: str | None = None,
+                   why: str | None = None) -> Manifest:
+    """Announce a booth, atomically (CLAUDE.md invariant 5).
+
+    Temp file + `os.replace`, because the CLI writes this in one process while
+    the browser reads it in another — a reader must never see a half-written
+    document. The temp file is itself a dotfile, so no listing, gallery or zip
+    can see it mid-write either.
+
+    OMITTED MEANS UNCHANGED; `""` MEANS CLEAR. `title` and `why` default to
+    None, not to the empty string, because the ordinary sequence is
+    `booth new x --why "..."` and then `booth add x out/*.png` — and while
+    omission meant empty, that second command silently erased the sentence the
+    first one existed to record. Two arms of the contract panel predicted it
+    from the wording alone; every test written for this module passed `--why`
+    on both calls and so could not see it.
+
+    RE-ANNOUNCING PRESERVES `created` (INV-3). It is when the booth APPEARED,
+    and saying something more about it later is not a second appearance. A
+    `created` that cannot be read back is replaced rather than guessed at: a
+    stamp that is silently wrong is worse than one that is silently new.
+
+    An empty `handle` becomes `SERVICE_HANDLE` rather than being refused — a
+    manifest with no handle does not read back at all, and an unreadable file is
+    the worse outcome. Unreachable from the CLI, whose fallback chain always
+    yields something; callers of this function directly should pass a real one.
+    """
+    booth = Path(booth)
+    booth.mkdir(parents=True, exist_ok=True)
+    prior = read_manifest(booth)
+    usable = prior if prior and not prior.error else None
+    created = usable.created if usable and usable.created else _now()
+
+    record = Manifest(
+        handle=_one_line(handle, HANDLE_MAX) or SERVICE_HANDLE,
+        title=(_one_line(title, TITLE_MAX) if title is not None
+               else (usable.title if usable else "")) or _one_line(booth.name, TITLE_MAX),
+        why=(_one_line(why, WHY_MAX) if why is not None
+             else (usable.why if usable else "")),
+        created=created,
+    )
+    path = booth / MANIFEST_FILE
+    doc = {"handle": record.handle, "title": record.title,
+           "why": record.why, "created": record.created}
+
+    # A write that changes nothing is not activity and must not reset the
+    # booth's TTL — the rule marks learned in v0.2.0, applied here because
+    # `booth link` re-announces the standing board on EVERY post to it.
+    if prior is not None and not prior.error and _as_doc(prior) == doc:
+        return record
+
+    # NOTHING THAT COULD NOT BE READ IS DESTROYED. Reads stay lenient, writes
+    # go strict, damaged bytes stay on disk — the doctrine marks made explicit
+    # in v0.2.1, which this write path contradicted by replacing them outright.
+    # A file that fails on ONE field still holds the others, and a `why` the
+    # re-announcer never kept anywhere is exactly what went missing.
+    #
+    # QUARANTINED rather than REFUSED, which is where this diverges from marks:
+    # refusing would fail `booth add` and lose the files it was mid-way through
+    # copying, and a booth's own description is restatable in a way the
+    # operator's judgment is not.
+    if prior is not None and prior.error:
+        try:
+            os.replace(path, booth / QUARANTINE_FILE)
+        except OSError:
+            pass                      # nothing to preserve beats failing the write
+
+    tmp = _temp_path(booth)
+    try:
+        tmp.write_text(
+            json.dumps(doc, ensure_ascii=False, indent=2) + "\n",
+            encoding="utf-8",
+        )
+        os.replace(tmp, path)
+    except BaseException:
+        # A leaked temp is worse here than it would be with a fixed name: the
+        # unique suffix means nothing ever overwrites it, and it is not a
+        # `.lock`, so `_newest_mtime` counts it and it keeps a dead booth alive
+        # forever. Cleaning up is the price of the uniqueness.
+        tmp.unlink(missing_ok=True)
+        raise
+    return record

booth/marks.py

@@ -42,6 +42,7 @@ from __future__ import annotations
 import fcntl
 import json
 import os
+import stat as statmod
 from dataclasses import asdict, dataclass, field
 from datetime import datetime
 from pathlib import Path
@@ -73,6 +74,14 @@ class MarksCorrupt(RuntimeError):
     """
 
 
+# A booth's whole judgment lives in one document, so this is generous — a
+# 270-item booth flagged throughout, with notes, is far under it. What it rules
+# out is the case that is not marks at all: an unbounded read raises MemoryError
+# and a deeply nested one raises RecursionError out of `json.loads`, neither of
+# which is an OSError or a ValueError, and `list_booths` calls the reader once
+# per booth on every index load. Bounded by `stat`, before the bytes are read.
+MARKS_MAX_BYTES = 4 * 1024 * 1024
+
 MARKS_FILE = ".marks.json"
 MARKS_LOCK = ".marks.lock"
 SCHEMA_VERSION = 1
@@ -126,7 +135,17 @@ class Mark:
 
 
 def now_stamp() -> str:
-    return datetime.now().astimezone().isoformat(timespec="seconds")
+    """ONE stamp format across every writer in this module.
+
+    MICROSECONDS, matching `import_legacy_asks`. They diverged when the
+    importer was moved to sub-second precision to stop same-second sidecars
+    re-sorting — and the divergence opened a fresh ordering bug in the other
+    direction, because `-` (0x2D) sorts before `.` (0x2E): a whole-second stamp
+    lands ahead of ANY fractional stamp in the same second, so a later mark came
+    out before an earlier import. Marks sort on `(created, id)`; one format is
+    what makes that rule statable.
+    """
+    return datetime.now().astimezone().isoformat(timespec="microseconds")
 
 
 def _clean_text(text) -> str:
@@ -173,9 +192,17 @@ def _read_raw(booth: Path) -> list[dict]:
     for the same reason: a review surface that will not load is worse than one
     that has lost an annotation.
     """
+    path = Path(booth) / MARKS_FILE
     try:
-        raw = json.loads((Path(booth) / MARKS_FILE).read_text(encoding="utf-8"))
+        st = path.stat()
-    except (OSError, ValueError, UnicodeDecodeError):
+        # Regular-file first, then size. `st_size` is 0 for a FIFO and 0 for a
+        # symlink to /dev/zero, so both pass a byte cap and then `read_text`
+        # either blocks with no EOF or allocates until the kernel intervenes.
+        # This loop runs over EVERY booth on every index load.
+        if not statmod.S_ISREG(st.st_mode) or st.st_size > MARKS_MAX_BYTES:
+            return []
+        raw = json.loads(path.read_text(encoding="utf-8"))
+    except (OSError, ValueError, UnicodeDecodeError, RecursionError, MemoryError):
         return []
     if not isinstance(raw, dict):
         return []
@@ -190,26 +217,51 @@ def _fingerprint(entries: list[dict]) -> str:
     return json.dumps(entries, sort_keys=True, ensure_ascii=False)
 
 
-def _read_raw_strict(booth: Path) -> list[dict]:
+def _read_raw_strict(booth: Path, *, blank_is_corrupt: bool = False) -> list[dict]:
     """Like `_read_raw`, but RAISES `MarksCorrupt` on a file it cannot parse.
 
     Absent, empty and valid-but-empty are all "no marks yet" and are fine — the
     distinction that matters is bytes-present-but-unreadable, because that is the
     case where writing would destroy something.
+
+    `blank_is_corrupt` is the DELETE path's reading of a present-but-whitespace
+    file, and only the delete path's: this writer never produces a blank marks
+    document, so a blank one that exists is something that went wrong, and
+    `rmtree` is not the response to that. The write path keeps the lenient
+    reading — a blank file is safe to overwrite, which is the question
+    `_Locked` is asking. A VALID document with an empty `marks` list is not
+    blank and never holds: that is what deleting the last mark leaves behind,
+    and it must stay sweepable.
     """
     path = Path(booth) / MARKS_FILE
+    try:
+        st = path.stat()
+    except FileNotFoundError:
+        return []
+    except OSError as exc:
+        raise MarksCorrupt(f"{path} cannot be read: {exc}") from exc
+    # The strict half has to refuse everything the lenient half tolerates, or a
+    # file that reads as "no marks" gets replaced by a write that believed it.
+    if not statmod.S_ISREG(st.st_mode):
+        raise MarksCorrupt(f"{path} is not a regular file")
+    if st.st_size > MARKS_MAX_BYTES:
+        raise MarksCorrupt(
+            f"{path} is too large to be a marks document ({st.st_size} bytes)")
     try:
         text = path.read_text(encoding="utf-8")
     except FileNotFoundError:
         return []
-    except (OSError, UnicodeDecodeError) as exc:
+    except (OSError, UnicodeDecodeError, MemoryError) as exc:
         raise MarksCorrupt(f"{path} cannot be read: {exc}") from exc
     if not text.strip():
+        if blank_is_corrupt:
+            raise MarksCorrupt(f"{path} is present but holds no marks document")
         return []
     try:
         raw = json.loads(text)
-    except ValueError as exc:
+    except (ValueError, RecursionError, MemoryError) as exc:
-        raise MarksCorrupt(f"{path} is not valid JSON: {exc}") from exc
+        raise MarksCorrupt(
+            f"{path} is not valid JSON: {type(exc).__name__}") from exc
     if not isinstance(raw, dict) or not isinstance(raw.get("marks"), list):
         raise MarksCorrupt(f"{path} is not a marks document")
     entries = [e for e in raw["marks"] if isinstance(e, dict) and isinstance(e.get("id"), str)]
@@ -218,14 +270,41 @@ def _read_raw_strict(booth: Path) -> list[dict]:
     return entries
 
 
+def read_error(booth: Path) -> str | None:
+    """Why this booth's marks cannot be read, or None if they can.
+
+    `marks_for` is lenient on purpose — a review page that will not load is
+    worse than one missing an annotation — and that leniency turns an
+    unreadable file into "no marks". For a BROWSER that is the right trade. For
+    the CLI it is not: a session that asked a question and is told "no such
+    pick" will conclude the question was never posted, when in fact the file
+    holding it is damaged. A machine consumer can act on the difference, so it
+    gets to ask.
+    """
+    try:
+        _read_raw_strict(booth)
+    except MarksCorrupt as exc:
+        return str(exc)
+    return None
+
+
 def _write_raw(booth: Path, entries: list[dict]) -> None:
     """Atomic replace, so a reader never sees a half-written document and a
     crash mid-write cannot truncate the file into a shorter — and therefore
     quieter — set of marks."""
     path = Path(booth) / MARKS_FILE
     doc = {"version": SCHEMA_VERSION, "marks": entries}
+    body = json.dumps(doc, ensure_ascii=False, indent=2) + "\n"
+    # The read bound is on the STORED bytes and `indent=2` grows them, so a
+    # document that fits in memory can land over the limit on disk and then read
+    # back as no marks at all. Refuse loudly instead: a write that fails is
+    # recoverable, a file that silently empties is not.
+    if len(body.encode("utf-8")) > MARKS_MAX_BYTES:
+        raise MarksCorrupt(
+            f"{path} would be larger than this version can read back "
+            f"({len(body.encode('utf-8'))} bytes)")
     tmp = path.with_suffix(path.suffix + ".tmp")
-    tmp.write_text(json.dumps(doc, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
+    tmp.write_text(body, encoding="utf-8")
     os.replace(tmp, path)
 
 
@@ -249,12 +328,48 @@ class _Locked:
         self.booth.mkdir(parents=True, exist_ok=True)
         lock = self.booth / MARKS_LOCK
         # `touch(exist_ok=True)` on an EXISTING file bumps its mtime, and a
-        # booth's TTL is measured from its newest mtime including dotfiles — so
+        # booth's TTL is measured from its newest mtime — so an unconditional
-        # an unconditional touch would keep a booth alive just for being read
+        # touch would keep a booth alive just for being read through a write
-        # through a write path. Create it only when it is not there.
+        # path. Create it only when it is not there.
+        #
+        # ONCE CREATED, THE LOCK FILE IS NEVER REMOVED (see __exit__).
         if not lock.exists():
+            # Creating a directory entry bumps the DIRECTORY's mtime, which is
+            # what `_newest_mtime` reads. An earlier version put the clock back
+            # with `os.utime` — which closed the bug and opened a race: the
+            # restore ran before the flock, so anything landing in the window
+            # between the stat and the utime had its bump rolled backward. An
+            # `rsync -a` batch is the case that bites, because it PRESERVES
+            # source mtimes and so has only the directory's freshness to look
+            # alive by. It could also raise OSError on a read-only directory
+            # and take the route down with it.
+            #
+            # THE RESTORE STAYS, and the honest reason is that the alternative
+            # was worse. Ignoring a booth directory's own mtime whenever the
+            # booth holds anything would close the race outright — and would
+            # also silently retire the documented behaviour that RELEASING a
+            # kept board resets its clock, which the CLI header, the README and
+            # a deliberate test all pin. That is a TTL doctrine change, not a
+            # bug fix, and it does not belong in one.
+            #
+            # ⚠ RESIDUAL RACE, stated rather than papered over: between the stat
+            # and the utime, another writer's directory-entry change can be
+            # rolled backward. The case that bites is an `rsync -a` batch, which
+            # preserves source mtimes and so has only the directory's freshness
+            # to look alive by. The window is the two syscalls below and the
+            # booth must also be one being written to at that instant.
+            #
+            # The concrete half IS fixed: a failing utime (read-only directory,
+            # a booth whose owner we are not) used to escape and take the whole
+            # route down with a 500. Not putting the clock back is a cost this
+            # module can absorb; not answering the request is not.
+            before = self.booth.stat()
             lock.touch()
             self._made_lock = True
+            try:
+                os.utime(self.booth, (before.st_atime, before.st_mtime))
+            except OSError:
+                pass
         self._lf = lock.open("r+")
         fcntl.flock(self._lf, fcntl.LOCK_EX)
         try:
@@ -264,8 +379,6 @@ class _Locked:
             fcntl.flock(self._lf, fcntl.LOCK_UN)
             self._lf.close()
             self._lf = None
-            if self._made_lock:
-                lock.unlink(missing_ok=True)
             raise
         self._before = _fingerprint(self.entries)
         return self
@@ -284,10 +397,17 @@ class _Locked:
             # would otherwise keep a dead booth alive forever.
             if exc_type is None and _fingerprint(self.entries) != self._before:
                 _write_raw(self.booth, self.entries)
-            elif self._made_lock and not (self.booth / MARKS_FILE).exists():
+            # THE LOCK FILE IS NEVER UNLINKED. It used to be, on the no-op path,
-                # Nothing was written and this booth had no marks before: do not
+            # so a booth that had never been marked was left exactly as it was
-                # leave a lock file behind as the only trace of a no-op.
+            # found. That tidiness cost mutual exclusion outright: `flock` binds
-                (self.booth / MARKS_LOCK).unlink(missing_ok=True)
+            # to an INODE, so unlinking the lock while a second writer is blocked
+            # on it leaves that writer holding an exclusive lock on a deleted
+            # file, and the NEXT writer creates a fresh lock and takes it at
+            # once. Two processes then run the read-modify-write concurrently,
+            # the later `os.replace` drops the earlier one's mark, and both of
+            # them obeyed the protocol. A zero-byte dotfile is the cheaper
+            # thing to leave behind — `booth_items` skips it, the zip skips it,
+            # and `_newest_mtime` exempts it so it cannot hold a booth open.
         finally:
             fcntl.flock(lf, fcntl.LOCK_UN)
             lf.close()
@@ -301,6 +421,25 @@ class _Locked:
 # ---- read -------------------------------------------------------------------
 
 
+def _entry_type_error(entry: dict) -> str | None:
+    """The stored scalars this module refuses to guess at.
+
+    `_clean_text` did `(text or "").replace(...)` and `marks_for` sorts on
+    `(created, id)` — so a stored `text` that is a dict, or a `created` that is a
+    number, raised AttributeError or TypeError out of the READ path. That is not
+    a marks bug, it is an INDEX bug: `list_booths` reads every booth's marks on
+    every page load and `/healthz` does the same, so one hand-edited or
+    foreign-written file took down the front page for every booth on the
+    service. A wrong type is a broken mark, and this module already knows how to
+    render one of those.
+    """
+    for name in ("created", "by", "text", "error"):
+        value = entry.get(name)
+        if value is not None and not isinstance(value, str):
+            return f"{name} is {type(value).__name__}, not a string"
+    return None
+
+
 def _hydrate(entry: dict) -> Mark:
     """One stored entry -> one Mark, declarations normalized.
 
@@ -312,6 +451,15 @@ def _hydrate(entry: dict) -> Mark:
     """
     mid = entry["id"]
     shape = entry.get("shape") if entry.get("shape") in SHAPES else NOTE
+    bad = _entry_type_error(entry)
+    if bad is not None:
+        # `created` is dropped rather than coerced, which sorts the entry to the
+        # TOP of the booth's marks: a mark nobody can read is the one that wants
+        # looking at, and burying it under 270 items' worth of notes is how it
+        # stays unnoticed. Deterministic, and stated — `("", id)` against
+        # `(created, id)`.
+        return Mark(id=mid, shape=shape, target=None, created="",
+                    error=f"unreadable mark: {bad}")
     target = entry.get("target")
     if not _valid_target(target):
         target = None
@@ -340,6 +488,40 @@ def _hydrate(entry: dict) -> Mark:
             norm = normalize_ask(decl, mid)
         except AskError as exc:
             return Mark(**base, declaration=decl, answer=answer, error=str(exc))
+        # THE ANSWER'S SHAPE IS VALIDATED HERE, at the ONE boundary every
+        # surface crosses — not at the three render sites that happen to draw
+        # it today, and not defensively in the template, which would hide that
+        # anything is wrong.
+        #
+        # `{"answer": {"answers": [], "notes": ""}}` is well-formed JSON with a
+        # wrong-shaped value. It passed `_entry_type_error`, passed the
+        # `isinstance(answer, dict)` check above, and `marks_for` and
+        # `hold_read` both reported the mark HEALTHY with no read error — and
+        # then `_ask_inline.html` did `a.answer.answers.get(q.key)`, Jinja asked
+        # a LIST for `.get`, and the gallery page and the marks page returned
+        # 500. Measured at 42ea67f, so it predates U3; U3 guarded only its own
+        # surface with `_safe_fragments` and left these two by scope.
+        #
+        # This is the v0.2.2 lesson finished rather than half-done. That outage
+        # was a file that could not be PARSED and the reader was made lenient;
+        # this one parses perfectly and breaks one layer further in, at render,
+        # where no leniency exists. `read_error` was answering a narrower
+        # question than every caller assumed.
+        #
+        # ONLY the multi case is checked, because only the multi case indexes:
+        # a single-question pick's answer IS the record, with no `answers` key
+        # to get wrong. Requiring one unconditionally would break every single
+        # pick, which is the direction a too-eager guard fails in.
+        if norm["multi"] and isinstance(answer, dict) and \
+                not isinstance(answer.get("answers"), dict):
+            return Mark(**base, declaration=decl, answer=None,
+                        prompt=norm["prompt"], title=norm["title"],
+                        multi=norm["multi"], questions=norm["questions"],
+                        options=norm.get("options", []),
+                        notes_enabled=norm["notes"], notes_label=norm["notes_label"],
+                        error="this pick's answer is stored in a shape the page "
+                              "cannot render; the answer was dropped and the "
+                              "question is unanswered")
         return Mark(
             **base,
             declaration=decl,
@@ -359,11 +541,26 @@ def _hydrate(entry: dict) -> Mark:
     return Mark(**base, text=_clean_text(entry.get("text")))
 
 
+def _hydrate_safe(entry: dict) -> Mark:
+    """`_hydrate`, with the promise that it cannot raise.
+
+    `_entry_type_error` covers the shapes we know how to name; this is the
+    backstop for the ones we do not, and it exists because of WHERE this runs.
+    One unreadable mark must cost that mark, never the page — and on the index
+    it is not even that booth's page, it is all of them.
+    """
+    try:
+        return _hydrate(entry)
+    except Exception as exc:                       # noqa: BLE001 - deliberate
+        return Mark(id=str(entry.get("id", "")), shape=NOTE, target=None,
+                    created="", error=f"unreadable mark: {exc}")
+
+
 def marks_for(booth: Path) -> list[Mark]:
     """Every mark in a booth, oldest first, declarations normalized and answers
     folded in. ONE file read — which is the whole point of the storage shape."""
     entries = _read_raw(booth)
-    marks = [_hydrate(e) for e in entries]
+    marks = [_hydrate_safe(e) for e in entries]
     # (created, id) rather than created alone: two marks written in the same
     # second would otherwise order by however json listed them.
     marks.sort(key=lambda m: (m.created, m.id))
@@ -393,6 +590,34 @@ def open_marks(marks: Sequence[Mark]) -> list[Mark]:
     return [m for m in marks if _is_open(m)]
 
 
+def hold_read(booth: Path) -> tuple[list[Mark], str | None]:
+    """ONE read of `.marks.json`, answering both questions the LIFETIME rule asks:
+    what is still open, and whether the file could be read at all.
+
+    U4 decides whether a booth may be SWEPT from those two facts. Asking them
+    with two calls — `marks_for` then `read_error` — reads the file twice, and
+    two reads of one file are not one read of one state: a write or a repair
+    landing between them yields a pair that never described the booth at any
+    instant. The losing pair is `([], None)` — no marks, no error — which is
+    exactly the one that deletes. Cross-frontier review (2026-09-22) found it;
+    that is why this exists rather than the obvious two calls.
+
+    When the file reads clean the marks are byte-identical to `marks_for`'s:
+    `_read_raw_strict` raises rather than dropping an entry, so a non-raising
+    strict read returns the same entries the lenient read would, hydrated and
+    sorted the same way. The caller can therefore use this ONE read for the
+    display too, and fall back to `marks_for` only on the error path, where
+    leniency is the point.
+    """
+    try:
+        entries = _read_raw_strict(booth, blank_is_corrupt=True)
+    except MarksCorrupt as exc:
+        return [], str(exc)
+    marks = [_hydrate_safe(e) for e in entries]
+    marks.sort(key=lambda m: (m.created, m.id))
+    return marks, None
+
+
 def marks_for_target(marks: Sequence[Mark], rel: str | None) -> list[Mark]:
     """The marks attached to one item, or to the booth itself for None."""
     return [m for m in marks if m.target == rel]
@@ -597,7 +822,8 @@ def import_legacy_asks(booth: Path) -> list[Mark]:
             continue
         try:
             decl = json.loads(p.read_text(encoding="utf-8"))
-        except (OSError, ValueError, UnicodeDecodeError) as exc:
+        except (OSError, ValueError, UnicodeDecodeError,
+                RecursionError, MemoryError) as exc:
             found.append((mtime, stem, None, f"unreadable ask: {exc}"))
             continue
         if not isinstance(decl, dict):
@@ -619,7 +845,8 @@ def import_legacy_asks(booth: Path) -> list[Mark]:
                 loaded = json.loads(ap.read_text(encoding="utf-8"))
                 if isinstance(loaded, dict):
                     answer = loaded
-            except (OSError, ValueError, UnicodeDecodeError):
+            except (OSError, ValueError, UnicodeDecodeError,
+                    RecursionError, MemoryError):
                 pass
 
             prior = by_id.get(stem)
@@ -642,7 +869,17 @@ def import_legacy_asks(booth: Path) -> list[Mark]:
                 "id": stem,
                 "shape": PICK,
                 "target": None,
-                "created": datetime.fromtimestamp(mtime).astimezone().isoformat(timespec="seconds"),
+                # MICROSECONDS, not seconds. `found` is ordered by fractional
+                # mtime and `marks_for` re-sorts on this string, so truncating
+                # to the whole second threw away the only thing distinguishing
+                # two sidecars written in the same second — and the `(created,
+                # id)` tie-break then silently re-sorted them alphabetically,
+                # reversing the order the importer had just established. The
+                # ROADMAP states this import's order is `(mtime, name)`; an
+                # order that is stated and not kept is worse than one never
+                # claimed.
+                "created": datetime.fromtimestamp(mtime).astimezone().isoformat(
+                    timespec="microseconds"),
                 "declaration": decl,
                 "answer": answer,
             }
@@ -653,4 +890,6 @@ def import_legacy_asks(booth: Path) -> list[Mark]:
 
     # Hydrated AFTER the lock so a broken declaration surfaces as `error` here
     # exactly as it does on a normal read, rather than through a second path.
-    return [_hydrate(e) for e in created]
+    # `_hydrate_safe`, not `_hydrate`: this is the one path that reads entries
+    # it did not write, and it was the one without the guard.
+    return [_hydrate_safe(e) for e in created]

booth/static/embed.js

@@ -0,0 +1,351 @@
+/* The Booth - the declared embed seam (U3).
+ *
+ * A booth that ships its own index.html is served verbatim. This script is how
+ * the Booth's chrome gets onto that page WITHOUT the Booth reaching into it:
+ * the report carries one line,
+ *
+ *     <script src="/_booth/embed.js" defer></script>
+ *
+ * and everything below mounts through real DOM APIs. It replaced ten regular
+ * expressions applied to author HTML - six hunting for a place to hang a
+ * favicon and a chip, four substituting rendered markup into the author's own
+ * tags. A page that declares this line is now served exactly as written.
+ *
+ * WHAT THIS SCRIPT DOES NOT DECIDE: what a mark says, whether it is still open,
+ * or what order marks come in. Every fragment below is rendered server-side by
+ * the same Jinja macros the gallery page uses, and `open` is computed by
+ * `open_marks`. Two renderers of one truth is the bug INV-1 exists to stop -
+ * the zoom view once re-derived an item and lost its captions doing it.
+ *
+ * Served from a read taken ONCE at app startup. Editing this file does nothing
+ * until `systemctl --user restart booth.service`, exactly like the templates,
+ * and for the same reason: on 2026-09-21 a hot-reloading template put 19 of 25
+ * booths at 500 against Python that had never heard of the context it wanted.
+ */
+(function () {
+  "use strict";
+  if (window.__boothEmbed) return;            // declared AND appended: mount once
+  window.__boothEmbed = true;
+
+  var CSS = [
+    /* ---- the way home, and the open-asks jump ---- */
+    ".booth-nav-home,.booth-nav-asks{position:fixed;top:0;z-index:2147483647;",
+    "display:inline-block;margin:.6rem;padding:.34rem .72rem;border-radius:8px;",
+    "text-decoration:none;letter-spacing:.01em;box-shadow:0 2px 10px rgba(0,0,0,.35)}",
+    /* top-right: a top-left chip clips the page title on left-aligned report
+       layouts, and this matches the zoom view's back affordance. */
+    ".booth-nav-home{right:0;font:600 13px/1.25 ui-sans-serif,system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;",
+    "color:#dfe7ef;background:rgba(20,23,32,.82);border:1px solid rgba(66,220,209,.35);",
+    "-webkit-backdrop-filter:blur(6px);backdrop-filter:blur(6px);transition:background .18s,border-color .18s}",
+    ".booth-nav-home:hover{background:rgba(28,33,46,.95);border-color:rgba(66,220,209,.75)}",
+    ".booth-nav-asks{right:7.2rem;font:700 13px/1.25 ui-sans-serif,system-ui,-apple-system,'Segoe UI',Roboto,sans-serif;",
+    "color:#171a23;background:#ffe14e;border:1px solid #ffe14e;transition:filter .18s}",
+    ".booth-nav-asks:hover{filter:brightness(1.08)}",
+    "@media print{.booth-nav-home,.booth-nav-asks{display:none}}",
+    /* ---- ask fragments. Self-contained: the host page carries its own CSS and
+       nothing here may inherit from it, so the palette adapts via
+       prefers-color-scheme rather than borrowing. ---- */
+    ".bk-ask{margin:1.1rem 0;padding:.85rem .95rem;border:1px solid rgba(128,140,160,.34);",
+    "border-top:2px solid #e0b93c;border-radius:9px;background:rgba(128,140,160,.07);",
+    "font:15px/1.5 ui-sans-serif,system-ui,-apple-system,'Segoe UI',Roboto,sans-serif}",
+    ".bk-ask.bk-done{border-top-color:#3fae6a}",
+    ".bk-ask.bk-skip{border-top-color:#6f7c8c}",
+    ".bk-ask.bk-skip .bk-ask-tag{color:#8a97a6}",
+    ".bk-ask-tag{display:block;margin-bottom:.5rem;font:700 10px/1 ui-monospace,SFMono-Regular,Menlo,monospace;",
+    "letter-spacing:.12em;text-transform:uppercase;color:#c9a227}",
+    ".bk-ask.bk-done .bk-ask-tag{color:#3fae6a}",
+    ".bk-ask-title{margin:0 0 .15rem;font-size:.72rem;letter-spacing:.07em;text-transform:uppercase;opacity:.62}",
+    ".bk-ask-prompt{margin:0 0 .6rem;font-weight:600}",
+    ".bk-ask-opts{display:flex;flex-direction:column;gap:.3rem}",
+    ".bk-ask-opt{display:flex;align-items:flex-start;gap:.55rem;padding:.45rem .6rem;cursor:pointer;",
+    "border:1px solid rgba(128,140,160,.3);border-radius:6px;background:rgba(128,140,160,.06)}",
+    ".bk-ask-opt:hover{border-color:rgba(128,140,160,.62)}",
+    ".bk-ask-opt:has(input:checked){border-color:#2fa8a0;background:rgba(47,168,160,.13)}",
+    ".bk-ask-opt input{margin:.25rem 0 0;flex:0 0 auto;accent-color:#2fa8a0}",
+    ".bk-ask-lab{display:flex;flex-direction:column;gap:.1rem;min-width:0}",
+    ".bk-ask-det{font-size:.8rem;opacity:.68}",
+    ".bk-ask-notes{display:block;width:100%;box-sizing:border-box;margin:.6rem 0 0;padding:.5rem .6rem;",
+    "font:inherit;font-size:.9rem;color:inherit;background:rgba(128,140,160,.09);",
+    "border:1px solid rgba(128,140,160,.34);border-radius:6px;resize:vertical}",
+    ".bk-ask-go{margin-top:.7rem;cursor:pointer;font:700 12px/1 ui-monospace,SFMono-Regular,Menlo,monospace;",
+    "letter-spacing:.06em;padding:.6rem 1.1rem;border-radius:6px;border:1px solid #2fa8a0;",
+    "background:#2fa8a0;color:#08131a}",
+    ".bk-ask-go:hover{filter:brightness(1.09)}",
+    ".bk-ask-was{margin:.15rem 0 .55rem;font-size:.84rem;opacity:.8}",
+    ".bk-ask-was b{opacity:1}",
+    ".bk-ask-err{color:#d6452a;font-size:.86rem}",
+    "@media (prefers-color-scheme: light){.bk-ask-tag{color:#8a6d10}.bk-ask-go{color:#fff}}",
+    "@media print{.bk-ask{break-inside:avoid}}"
+  ].join("");
+
+  /* The anchor attributes. `data-booth-mark` is canonical - U2 made an ask one
+     shape of mark - and `data-booth-ask` is kept because two of the four live
+     verbatim booths spell it that way, in the operator's own reports. */
+  var MAIN_SEL = "[data-booth-mark],[data-booth-ask]";
+  var SUBMIT_SEL = "[data-booth-mark-submit],[data-booth-ask-submit]";
+  var WHOLE = " whole";   // the set member meaning "the whole ask landed here"
+
+  function attr(el, a, b) {
+    var v = el.getAttribute(a);
+    return v === null ? el.getAttribute(b) : v;
+  }
+
+  /* "batch:r1" -> ["batch", "r1"]; "batch" -> ["batch", null]. Split on the
+     FIRST colon: a question key cannot contain one (asks._KEY_RE) and neither
+     can a pick id (asks.valid_stem), so this is unambiguous for everything the
+     payload carries. A flag's id IS `flag:<target>`, which is why the payload
+     carries picks only. */
+  function splitSpec(spec) {
+    var s = (spec || "").trim();
+    var i = s.indexOf(":");
+    if (i < 0) return [s, null];
+    return [s.slice(0, i).trim(), s.slice(i + 1).trim() || null];
+  }
+
+  function boothName() {
+    var tag = document.querySelector("script[data-booth]");
+    if (tag) return tag.getAttribute("data-booth");
+    var parts = location.pathname.split("/");   // ["", "b", "<name>", ...]
+    if (parts.length < 3 || parts[1] !== "b" || !parts[2]) return null;
+    try { return decodeURIComponent(parts[2]); } catch (e) { return parts[2]; }
+  }
+
+  function mount(el, html) {
+    /* beforeend, NOT replaceWith: the author's element and its contents survive
+       and the fragment lands inside it. `<div class="ask" data-booth-ask="...">
+       <h3>heading</h3>` is live markup today, and the regex it replaced ate
+       both the wrapper class and the heading's framing.
+
+       Returns the elements it actually inserted. The chip needs to jump to a
+       fragment WE mounted, not to whatever the document happens to have with a
+       matching id — see chipTarget. */
+    var before = el.children.length;
+    el.insertAdjacentHTML("beforeend", html);
+    return Array.prototype.slice.call(el.children, before);
+  }
+
+  function styles() {
+    if (document.getElementById("booth-embed-css")) return;
+    var st = document.createElement("style");
+    st.id = "booth-embed-css";
+    st.textContent = CSS;
+    (document.head || document.documentElement).appendChild(st);
+  }
+
+  function favicon(href) {
+    /* The question `_ICON_RE` and its three head-seam siblings were asking of
+       raw text. Same question, asked of a parsed document. */
+    if (!href || document.querySelector('link[rel~="icon"]')) return;
+    var link = document.createElement("link");
+    link.rel = "icon";
+    link.href = href;
+    (document.head || document.documentElement).appendChild(link);
+  }
+
+  function homeChip(href) {
+    var a = document.createElement("a");
+    a.className = "booth-nav-home";
+    a.href = href || "/";
+    a.setAttribute("aria-label", "back to all booths");
+    a.textContent = "‹ all booths";
+    document.body.appendChild(a);
+  }
+
+  function chipTarget(mounted, markId) {
+    /* The earliest IN DOCUMENT ORDER of the elements WE mounted for this mark.
+       Not an id-prefix search over the whole document: a panel pointed out that
+       an author's own `<section id="bk-ask-winner-background">` satisfies any
+       prefix rule — hyphen boundary included — and would hijack the jump. Only
+       elements this script inserted are candidates, which is the identity the
+       deleted `bk-ask-<id>-top` anchor used to guarantee. */
+    var mine = mounted[markId] || [];
+    var first = null;
+    for (var i = 0; i < mine.length; i++) {
+      var el = mine[i];
+      if (!el.id || !document.contains(el)) continue;
+      if (first === null ||
+          (first.compareDocumentPosition(el) & Node.DOCUMENT_POSITION_PRECEDING)) {
+        first = el;
+      }
+    }
+    return first;
+  }
+
+  function asksChip(openIds, mounted) {
+    if (!openIds.length) return;
+    /* A JUMP LINK, not a way out to another page: on a long report the question
+       can be well below the fold and "there is a question waiting" still has to
+       be visible at first paint. */
+    var first = chipTarget(mounted, openIds[0]);
+    var a = document.createElement("a");
+    a.className = "booth-nav-asks";
+    a.href = first ? "#" + first.id : "/b/" + encodeURIComponent(boothName() || "") + "/marks";
+    a.textContent = "? " + openIds.length + " open ask" + (openIds.length === 1 ? "" : "s");
+    document.body.appendChild(a);
+  }
+
+  function reassociate() {
+    /* SCOPED TO OUR OWN FRAGMENTS (`.bk-ask [form]`), deliberately: the Booth
+       does not rewrite attributes on elements the author wrote, even to help.
+       A control bound to its <form> by the HTML5 `form=` attribute resolves its
+       form owner when it is inserted. The fragments go in in VISUAL order, so a
+       question can land before the submit block that carries the <form>.
+       Chromium 151 re-resolves this correctly - measured 2026-09-22, N=3 per
+       condition, with a form-first positive control and a points-at-nothing
+       negative control. The sensitivity floor of that probe is ONE ENGINE, and
+       the failure it would hide is a form the operator fills in whose controls
+       reach no form at all, so the button does nothing and nothing is saved.
+       Three lines, so the engine stops mattering. */
+    var bound = document.querySelectorAll(".bk-ask [form]");
+    for (var i = 0; i < bound.length; i++) {
+      var v = bound[i].getAttribute("form");
+      bound[i].removeAttribute("form");
+      bound[i].setAttribute("form", v);
+    }
+  }
+
+  function hasForm(markId) {
+    /* `form_id` in booth/app.py builds the same string. Kept in step by the
+       fragments themselves: the submit macro emits exactly this id. */
+    return !!document.getElementById(
+      "bk-ask-form-" + markId.replace(/[^A-Za-z0-9_-]/g, "-"));
+  }
+
+  function place(marks) {
+    /* Object.create(null), NOT {} — three times, and it is not style.
+       A mark id and a question key are both `[A-Za-z0-9][A-Za-z0-9._-]*`
+       (asks.valid_stem, asks._KEY_RE), so `toString` and `constructor` are
+       legal in both. Against a plain object, an author writing
+       `data-booth-mark="toString"` — an anchor naming NO mark — gets
+       Object.prototype.toString back, passes the `if (!mark)` guard it was
+       supposed to fail, and throws on `mark.questions.length`. That aborts
+       `place` before the tail, so the page loses EVERY ask, from one typo in
+       the author's own markup. The `placed` set has the mirror bug: inherited
+       `got.constructor` reads as "already placed" and silently drops a real
+       question. Found by a cross-frontier code-review panel. */
+    var by = Object.create(null);
+    for (var i = 0; i < marks.length; i++) by[marks[i].id] = marks[i];
+    var placed = Object.create(null);      // id -> {key or WHOLE: true}
+    var submitted = Object.create(null);
+    var mounted = Object.create(null);     // id -> [elements this script inserted]
+
+    function note(id, key) {
+      if (!placed[id]) placed[id] = Object.create(null);
+      if (key !== undefined) placed[id][key] = true;
+    }
+
+    function record(id, els) {
+      if (!mounted[id]) mounted[id] = [];
+      for (var n = 0; n < els.length; n++) mounted[id].push(els[n]);
+    }
+
+    // 1. whole / per-question anchors, in DOCUMENT ORDER.
+    var anchors = document.querySelectorAll(MAIN_SEL);
+    for (var a = 0; a < anchors.length; a++) {
+      var el = anchors[a];
+      var spec = splitSpec(attr(el, "data-booth-mark", "data-booth-ask"));
+      var mark = by[spec[0]];
+      if (!mark) continue;                 // a typo'd id is LEFT ALONE, not blanked
+      if (spec[1] === null) {
+        record(mark.id, mount(el, mark.whole));
+        note(mark.id, WHOLE);
+        if (hasForm(mark.id)) submitted[mark.id] = true;
+        continue;
+      }
+      var q = null;
+      for (var k = 0; k < mark.questions.length; k++) {
+        if (mark.questions[k].key === spec[1]) { q = mark.questions[k]; break; }
+      }
+      if (!q) continue;                    // names no question: also left alone
+      record(mark.id, mount(el, q.html));
+      note(mark.id, spec[1]);
+    }
+
+    // 2. explicit submit anchors.
+    var subs = document.querySelectorAll(SUBMIT_SEL);
+    for (var s = 0; s < subs.length; s++) {
+      var sel = subs[s];
+      var sid = splitSpec(attr(sel, "data-booth-mark-submit", "data-booth-ask-submit"))[0];
+      var sm = by[sid];
+      if (!sm) continue;
+      /* A BROKEN pick has no submit block — its `submit` is the empty string and
+         its diagnostic lives in `whole`. Mounting nothing here and then marking
+         it placed made the tail skip it, so the "broken ask" box never rendered
+         at the one surface built to show it. Leave the anchor alone, exactly as
+         an anchor naming no mark is left alone, and let the tail mount the
+         diagnostic. */
+      if (sm.error) continue;
+      record(sm.id, mount(sel, sm.submit));
+      note(sm.id);
+      /* ...and only count it submitted if the <form> SURVIVED. An author who
+         puts this anchor inside their own <form> loses ours: the HTML parser
+         drops a nested form element outright. Every control's `form=` would
+         then point at nothing, the tail would not add a fallback because we
+         said it was handled, and the operator would fill the whole thing in and
+         click a button that does nothing. */
+      if (hasForm(sm.id)) submitted[sm.id] = true;
+    }
+
+    // 3. the tail, in PAYLOAD order - `(created, id)`. An ask is never
+    //    invisible: an unmarked page gets the whole thing, and a partially
+    //    marked one gets every question the author did not place, because a
+    //    question the operator cannot see is a question he cannot answer, and
+    //    a submission with NOTHING picked is refused outright (400), so a page
+    //    showing two of four questions can strand a pick that looks answerable.
+    //    (A PARTIAL answer is accepted and recorded — that is deliberate.)
+    var holder = document.createElement("div");
+    for (var m = 0; m < marks.length; m++) {
+      var mk = marks[m];
+      var got = placed[mk.id];
+      var was = holder.children.length;
+      if (!got) {
+        holder.insertAdjacentHTML("beforeend", mk.whole);
+        record(mk.id, Array.prototype.slice.call(holder.children, was));
+        continue;
+      }
+      if (mk.error) continue;
+      if (!got[WHOLE]) {
+        for (var q2 = 0; q2 < mk.questions.length; q2++) {
+          var qq = mk.questions[q2];
+          if (!got[qq.key]) holder.insertAdjacentHTML("beforeend", qq.html);
+        }
+      }
+      if (!submitted[mk.id]) holder.insertAdjacentHTML("beforeend", mk.submit);
+      record(mk.id, Array.prototype.slice.call(holder.children, was));
+    }
+    var tail = document.createDocumentFragment();
+    while (holder.firstChild) tail.appendChild(holder.firstChild);
+    document.body.appendChild(tail);
+    return mounted;
+  }
+
+  function start() {
+    var name = boothName();
+    if (!name || !document.body) return;
+    styles();
+    // Mounted BEFORE the fetch and from a constant, so a failed or slow fetch
+    // still leaves the operator a way out. That is why the payload carries no
+    // `home` — a value on the wire that nothing reads is a second
+    // representation of one fact, waiting to disagree with the first.
+    homeChip("/");
+    fetch("/b/" + encodeURIComponent(name) + "/embed.json", { credentials: "same-origin" })
+      .then(function (r) { return r.ok ? r.json() : null; })
+      .then(function (data) {
+        if (!data) return;
+        favicon(data.favicon);
+        var mounted = place(data.marks || []);
+        reassociate();
+        asksChip(data.open || [], mounted);
+        document.dispatchEvent(new CustomEvent("booth:mounted", { detail: { booth: name } }));
+      })
+      .catch(function () { /* the report is the operator's; a failed fetch costs
+                              the chrome, never the page. */ });
+  }
+
+  /* The declared line carries `defer`, but an author may not copy it exactly. */
+  if (document.readyState === "loading") {
+    document.addEventListener("DOMContentLoaded", start);
+  } else {
+    start();
+  }
+})();

booth/templates/_ask_inline.html

@@ -1,56 +1,22 @@
-{# Self-contained ask fragments injected into a booth's VERBATIM index.html.
+{# Self-contained ask fragments for a booth's VERBATIM index.html.
 
    The page is served untouched and carries its own CSS, so nothing here may
-   inherit from base.html: every fragment ships its own scoped `.bk-ask-*`
+   inherit from base.html. Since U3 these fragments do not reach the page by
-   styles (emitted once, by `styles()`), and the palette adapts via
+   string substitution: they are rendered here, handed over
-   prefers-color-scheme rather than borrowing the host page's.
+   `/b/<name>/embed.json`, and MOUNTED INTO THE DOM by `/_booth/embed.js`. The
+   scoped `.bk-ask-*` styles live in that file alongside the code that needs
+   them, which is why this template no longer emits a `styles()` block.
+
+   These macros stay the ONE renderer of an ask fragment. embed.js places what
+   comes back and never builds one.
 
    Per-question fragments are wired to ONE form with the HTML5 `form=`
    attribute, so a four-voice report can put each radio group under its own
-   audio block and still submit all four picks in a single POST — which is what
+   audio block and still submit all four picks in a single POST -- which is what
    the multi-question ask requires. The <form> element itself is empty and
-   lives with the submit block. No JavaScript.
+   lives with the submit block.
 #}
 
-{% macro styles() %}
-<style>
-.bk-ask{margin:1.1rem 0;padding:.85rem .95rem;border:1px solid rgba(128,140,160,.34);
-  border-top:2px solid #e0b93c;border-radius:9px;background:rgba(128,140,160,.07);
-  font:15px/1.5 ui-sans-serif,system-ui,-apple-system,"Segoe UI",Roboto,sans-serif}
-.bk-ask.bk-done{border-top-color:#3fae6a}
-.bk-ask.bk-skip{border-top-color:#6f7c8c}
-.bk-ask.bk-skip .bk-ask-tag{color:#8a97a6}
-.bk-ask-tag{display:block;margin-bottom:.5rem;font:700 10px/1 ui-monospace,SFMono-Regular,Menlo,monospace;
-  letter-spacing:.12em;text-transform:uppercase;color:#c9a227}
-.bk-ask.bk-done .bk-ask-tag{color:#3fae6a}
-.bk-ask-title{margin:0 0 .15rem;font-size:.72rem;letter-spacing:.07em;text-transform:uppercase;opacity:.62}
-.bk-ask-prompt{margin:0 0 .6rem;font-weight:600}
-.bk-ask-opts{display:flex;flex-direction:column;gap:.3rem}
-.bk-ask-opt{display:flex;align-items:flex-start;gap:.55rem;padding:.45rem .6rem;cursor:pointer;
-  border:1px solid rgba(128,140,160,.3);border-radius:6px;background:rgba(128,140,160,.06)}
-.bk-ask-opt:hover{border-color:rgba(128,140,160,.62)}
-.bk-ask-opt:has(input:checked){border-color:#2fa8a0;background:rgba(47,168,160,.13)}
-.bk-ask-opt input{margin:.25rem 0 0;flex:0 0 auto;accent-color:#2fa8a0}
-.bk-ask-lab{display:flex;flex-direction:column;gap:.1rem;min-width:0}
-.bk-ask-det{font-size:.8rem;opacity:.68}
-.bk-ask-notes{display:block;width:100%;box-sizing:border-box;margin:.6rem 0 0;padding:.5rem .6rem;
-  font:inherit;font-size:.9rem;color:inherit;background:rgba(128,140,160,.09);
-  border:1px solid rgba(128,140,160,.34);border-radius:6px;resize:vertical}
-.bk-ask-go{margin-top:.7rem;cursor:pointer;font:700 12px/1 ui-monospace,SFMono-Regular,Menlo,monospace;
-  letter-spacing:.06em;padding:.6rem 1.1rem;border-radius:6px;border:1px solid #2fa8a0;
-  background:#2fa8a0;color:#08131a}
-.bk-ask-go:hover{filter:brightness(1.09)}
-.bk-ask-was{margin:.15rem 0 .55rem;font-size:.84rem;opacity:.8}
-.bk-ask-was b{opacity:1}
-.bk-ask-err{color:#d6452a;font-size:.86rem}
-@media (prefers-color-scheme: light){
-  .bk-ask-tag{color:#8a6d10}
-  .bk-ask-go{color:#fff}
-}
-@media print{.bk-ask{break-inside:avoid}}
-</style>
-{% endmacro %}
-
 {# One question's radio group, bound to the shared form by id. #}
 {% macro question(a, q, form_id, name_url, standalone=False) %}
 {% set field = 'choice.' ~ q.key if a.multi else 'choice' %}

booth/templates/_lifetime.html

@@ -0,0 +1,28 @@
+{# The lifetime line, defined ONCE and called from four surfaces: the index
+   card (both lanes), the booth header (both branches) and the marks page.
+
+   U4: a booth's lifetime is derived from its own state, and a booth that is
+   not counting down must always SAY WHY — an invisible rule that silently
+   stopped the clock would be strictly worse than the `.forever` boolean it
+   replaces, because that one was at least visible as a lane.
+
+   `hold` is the REASON, straight off `hold_reason()`, not a bool beside a
+   string that can disagree with it. Kept wins over a hold because a kept booth
+   is exempt either way, and showing two reasons for one EXEMPTION is the
+   two-representations-of-one-state trap.
+
+   Unreadable marks are the exception and ride along even on a kept board:
+   damaged judgment is not a second exemption, it is a thing somebody has to go
+   and fix, and the kept lane holds the durable boards — the ones where losing
+   the operator's marks costs most. #}
+{% macro lifetime(kept, hold, expires_in) -%}
+  {%- if kept -%}
+    kept{% if hold == "unreadable" %} · <span class="held held-broken" title="a mark in this booth cannot be read">marks unreadable</span>{% endif %}
+  {%- elif hold == "unreadable" -%}
+    <span class="held held-broken" title="a mark in this booth cannot be read, so the sweeper will not take it">held · marks unreadable</span>
+  {%- elif hold == "open" -%}
+    <span class="held" title="an unanswered question holds this booth open">held until answered</span>
+  {%- else -%}
+    expires in {{ expires_in|dur }}
+  {%- endif -%}
+{%- endmacro %}

booth/templates/_marks.html

@@ -13,11 +13,35 @@
    Works with JS off — plain form POST, every shape. An answered pick shows the
    recorded judgment and a collapsed "change" form, because the mark is the
    CURRENT judgment and not a log. #}
+{# A mark carrying `error` is sorted out FIRST, whatever shape it claims. A
+   pick keeps its own ⚠ broken rendering below (richer — it has a declaration to
+   show); a broken note would otherwise render as an empty <pre> with a withdraw
+   button, indistinguishable from a note the operator wrote and then cleared,
+   and a broken flag would link to a target that is not there. Unreadable state
+   is visible state — the rule `_hydrate` states for picks, applied to all
+   three. #}
+{% set broken = marks | selectattr('error') | rejectattr('shape', 'equalto', 'pick') | list %}
 {% set picks = marks | selectattr('shape', 'equalto', 'pick') | list %}
-{% set notes = marks | selectattr('shape', 'equalto', 'note') | list %}
+{% set notes = marks | selectattr('shape', 'equalto', 'note') | rejectattr('error') | list %}
-{% set flags = marks | selectattr('shape', 'equalto', 'flag') | list %}
+{% set flags = marks | selectattr('shape', 'equalto', 'flag') | rejectattr('error') | list %}
 <section class="marks">
 
+{% for a in broken %}
+  <article class="mark mark-note is-broken" id="mark-{{ a.id }}">
+    <header class="mark-head">
+      <span class="mark-state">⚠ broken</span>
+      <span class="mark-id"><code>{{ a.id }}</code></span>
+      <span class="board-spacer"></span>
+      <form class="mark-undo" method="post" action="/b/{{ name_url }}/unmark">
+        <input type="hidden" name="mark" value="{{ a.id }}">
+        {% if marks_page %}<input type="hidden" name="back" value="marks">{% endif %}
+        <button type="submit" class="mark-x" title="withdraw this mark">×</button>
+      </form>
+    </header>
+    <p class="mark-error">This mark could not be read: {{ a.error }}</p>
+  </article>
+{% endfor %}
+
 {% for a in picks %}
   <article class="mark mark-pick{% if a.answer and a.answer.complete %} is-answered{% elif a.answer %} is-partial{% elif a.error %} is-broken{% endif %}" id="mark-{{ a.id }}">
     <header class="mark-head">

booth/templates/_provenance.html

@@ -0,0 +1,17 @@
+{# THE ANNOUNCEMENT — who posted this booth and why. Defined ONCE and called
+   from both index lanes and the booth page header: the kept lane is a separate
+   block, and patching only the ephemeral one would leave the durable,
+   most-looked-at boards with exactly the defect this closes.
+
+   Four states, and `unannounced` is distinct from `unreadable` on purpose —
+   folding "cannot be read" into "never said" hides the one case somebody has to
+   go and fix. The classes are the test hooks; the words are for the operator. #}
+{% macro provenance(m) -%}
+  {% if m is none %}
+    <div class="prov prov-none">unannounced</div>
+  {% elif m.error %}
+    <div class="prov prov-broken" title="{{ m.error }}">unreadable</div>
+  {% else %}
+    <div class="prov"><span class="prov-who">{{ m.handle }}</span>{% if m.why %} · <span class="prov-why">{{ m.why }}</span>{% endif %}</div>
+  {% endif %}
+{%- endmacro %}

booth/templates/base.html

@@ -255,6 +255,22 @@
   .card .name:hover{text-decoration:none;color:var(--aus-bright-cyan)}
   .card .sub{color:var(--fg-3);font-size:.72rem;font-family:var(--font-mono);letter-spacing:.03em;margin-top:.3rem}
 
+  /* THE ANNOUNCEMENT — who posted this booth and why (U5). Same size and
+     rhythm as .sub above it, because it is the same class of information: a
+     second line of card metadata, not a heading. The handle carries the only
+     colour, so a scan down the index reads as a column of posters. */
+  .prov{margin-top:.28rem;font-size:.72rem;font-family:var(--font-mono);
+    letter-spacing:.03em;color:var(--fg-3);line-height:1.45;
+    overflow-wrap:anywhere}
+  .prov-who{color:var(--fg-2)}
+  .prov-why{color:var(--fg-3)}
+  /* Quiet on purpose. 26 booths arrived before this convention existed and
+     rsync keeps making more, so the marker has to be visible-if-you-look and
+     never a badge shouting 26 times. `unreadable` gets the warning tint
+     because, unlike `unannounced`, it is something somebody has to fix. */
+  .prov-none{color:var(--fg-muted);font-style:italic}
+  .prov-broken{color:var(--aus-bright-yellow,#e8c547);font-style:italic;cursor:help}
+
   .wipe{position:absolute;top:.5rem;right:.5rem;margin:0}
   /* ★ keep, mirroring .wipe on the other shoulder of the card. Same
      hover-to-reveal language as .release in the kept lane. */
@@ -314,6 +330,11 @@
      use for state), green check once answered; the accent is a TOP edge, per
      Australis, never a coloured left border. */
   .badge-mark{background:var(--aus-bright-yellow);color:var(--fg-on-accent)}
+  /* U4: the lifetime line's HELD states. Marked rather than styled into
+     invisibility — the whole safety argument for an unbounded hold is that
+     a booth which stopped counting down says so where the countdown was. */
+  .held{color:var(--aus-bright-yellow)}
+  .held-broken{color:var(--fg-3);text-decoration:underline dotted}
   .thumb .badge+.badge-mark{top:2.2rem}
   .marks{display:flex;flex-direction:column;gap:.9rem;margin:.2rem 0 1.4rem}
   .mark{border:1px solid var(--border-subtle);border-top:2px solid var(--aus-bright-yellow);
@@ -408,6 +429,13 @@
   .boothhead h1{margin:0;font-family:var(--font-display);font-weight:600;font-size:1.5rem;
     letter-spacing:-.01em;word-break:break-word;flex:1 1 auto;color:var(--fg-0)}
   .boothhead .sub{color:var(--fg-3);font-size:.74rem;font-family:var(--font-mono);letter-spacing:.06em}
+  /* Its own row under the title, not another chip in the flex line — a `why`
+     can run to WHY_MAX and would otherwise shove the zip link around. */
+  .boothhead .prov{flex:0 0 100%;margin-top:-.35rem}
+  /* The directory name beside a manifest title: quieter than the title, but
+     never absent — it is what the URL says and what "the third one" refers to. */
+  .h1-slug{font-family:var(--font-mono);font-size:.62em;font-weight:400;
+    letter-spacing:.06em;color:var(--fg-3);margin-left:.5rem;white-space:nowrap}
   .wipe-lg{position:static}
   /* red-outline danger button — legible on the dark canvas, fills on hover */
   .wipe-lg button{width:auto;height:auto;padding:.42rem .85rem;border-radius:var(--radius-md);
@@ -468,6 +496,49 @@
   .markdown-body table{border-collapse:collapse;display:block;overflow-x:auto}
   .markdown-body th,.markdown-body td{border:1px solid var(--rk-line,#252a35);padding:.4em .7em}
   .markdown-body img{max-width:100%}
+
+    /* U6 — the bench registry, on the standing board's page only. */
+    .benches{margin:1rem 0;border:1px solid var(--line,#2a2a2a);border-radius:6px;overflow:hidden}
+    .bench-head{display:flex;gap:.6rem;align-items:baseline;padding:.5rem .7rem;background:rgba(255,255,255,.03)}
+    .bench-title{font-weight:600}
+    .bench-note,.bench-empty{opacity:.6;font-size:.85em}
+    .bench-empty{padding:.6rem .7rem}
+    .bench-err{padding:.6rem .7rem;color:#f2b8b5;background:rgba(242,184,181,.08)}
+    .bench-row{display:flex;gap:.6rem;align-items:center;padding:.45rem .7rem;border-top:1px solid var(--line,#2a2a2a)}
+    .bench-row.is-retired{opacity:.5}
+    .bench-state{font-size:.7em;text-transform:uppercase;letter-spacing:.06em;padding:.1rem .4rem;border-radius:3px;background:rgba(255,255,255,.08)}
+    .bench-row.is-live .bench-state{background:rgba(120,200,140,.18)}
+    .bench-row.is-promoted .bench-state{background:rgba(130,170,240,.18)}
+    .bench-main{flex:1;min-width:0}
+    .bench-url{font-size:.78em;opacity:.55;overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
+    .bench-meta{display:flex;flex-direction:column;align-items:flex-end;font-size:.75em;opacity:.6}
+    .bench-acts{display:flex;gap:.3rem}
+    .bench-to,.bench-rm{font-size:.75em;padding:.15rem .4rem;cursor:pointer}
+    .bench-add{display:flex;gap:.4rem;padding:.5rem .7rem;border-top:1px solid var(--line,#2a2a2a)}
+    .bench-add input[type=url]{flex:2;min-width:0}
+    .bench-add input[type=text]{flex:1;min-width:0}
+    /* A board row whose booth has been swept. Marked, never auto-removed. */
+    .board-row.board-dead{opacity:.45}
+    .board-dead-tag{font-size:.9em;color:#f2b8b5;opacity:.9}
+      /* U7 — the rail, and the grid cursor. */
+    .rail{position:sticky;top:0;z-index:5;display:flex;gap:.5rem;align-items:baseline;
+          padding:.4rem .6rem;margin:.6rem 0;background:var(--bg,#111);
+          border-bottom:1px solid var(--line,#2a2a2a);flex-wrap:wrap}
+    .rail-total{font-weight:600}
+    .rail-f{font-size:.85em;padding:.1rem .45rem;border-radius:3px;text-decoration:none;
+            opacity:.65;border:1px solid transparent}
+    .rail-f:hover{opacity:1}
+    .rail-f.on{opacity:1;border-color:var(--line,#2a2a2a);background:rgba(255,255,255,.06)}
+    /* The group row. Wraps rather than scrolls: 16 groups is the live maximum
+       and a horizontal scroller hides half of them behind a gesture. */
+    .rail-groups{display:flex;gap:.35rem;flex-wrap:wrap;align-items:baseline;
+                 padding-left:.5rem;margin-left:.25rem;border-left:1px solid var(--line,#2a2a2a)}
+    .rail-g{font-size:.8em;padding:.1rem .4rem;border-radius:3px;text-decoration:none;
+            opacity:.6;border:1px solid transparent}
+    .rail-g:hover{opacity:1;border-color:var(--line,#2a2a2a)}
+    /* The jumped-to tile, so a fragment jump says where it landed. */
+    figure.item:target{outline:2px dashed #7aa2f7;outline-offset:3px}
+    figure.item.is-cursor{outline:2px solid #7aa2f7;outline-offset:2px}
   </style>
 </head>
 <body>

booth/templates/booth.html

@@ -1,4 +1,6 @@
 {% extends "base.html" %}
+{% from "_provenance.html" import provenance %}
+{% from "_lifetime.html" import lifetime %}
 {# The blur toggle, defined ONCE. There are three item branches in this file
    (doc / media / other) and the first cut of this feature patched only one of
    them, so docs rendered with no control at all. A macro makes "patched two of
@@ -55,9 +57,18 @@
 {% block content %}
 <div class="boothhead">
   <a class="back" href="/">‹ all booths</a>
+  {# The manifest's TITLE is the display name; the directory name stays visible
+     beside it because that is the identity the operator navigates by and refers
+     to positionally, and losing it would be losing the thing the URL says.
+     Index cards keep the directory name alone for the same reason. #}
+  {% if manifest and not manifest.error and manifest.title and manifest.title != name %}
+    <h1>{{ manifest.title }} <span class="h1-slug">{{ name }}</span></h1>
+  {% else %}
     <h1>{{ name }}</h1>
-  <span class="sub">{% if uploaded %}<span class="badge">⬆ pickup</span> {% endif %}{% if board %}{{ board|length }} link{{ '' if board|length == 1 else 's' }}{% if items %} · {{ items|length }} file{{ '' if items|length == 1 else 's' }}{% endif %}{% else %}{% if marks_open %}<span class="badge badge-mark">{{ marks_open }} open</span> · {% endif %}{{ items|length }} item{{ '' if items|length == 1 else 's' }} · expires in {{ expires_in|dur }}{% endif %}</span>
+  {% endif %}
+  <span class="sub">{% if uploaded %}<span class="badge">⬆ pickup</span> {% endif %}{% if board %}{{ board|length }} link{{ '' if board|length == 1 else 's' }}{% if items %} · {{ items|length }} file{{ '' if items|length == 1 else 's' }}{% endif %} · {{ lifetime(kept, hold, expires_in) }}{% else %}{% if marks_open %}<span class="badge badge-mark">{{ marks_open }} open</span> · {% endif %}{{ items|length }} item{{ '' if items|length == 1 else 's' }} · {{ lifetime(kept, hold, expires_in) }}{% endif %}</span>
   {% if items %}<a class="dl-link" href="/b/{{ name_url }}/?download=1" title="download this booth as a zip">⬇ zip</a>{% endif %}
+  {{ provenance(manifest) }}
     {# A durable multi-writer board gets no one-click wipe — same rule as the
        kept lane on the index. Remove rows with the per-row ×, or release the
        board from the index and wipe it from there. #}
@@ -94,10 +105,83 @@
    back to the flagged items. Always rendered on a gallery booth — the add-note
    field is a control, not a result, so it has to be there before the first
    mark exists. #}
-{% if not board %}
+{# `marks or not board`: the standing link board renders as a board rather than
+   a gallery, and the add-note control would be noise on it — but the
+   suppression was unconditional, so a pick declared on a booth that happens to
+   carry a links.md had no form to answer it and nothing said so. #}
+{% if marks or not board %}
   {% include "_marks.html" %}
 {% endif %}
 
+{# THE BENCH REGISTRY — BLOCK LEVEL, and that placement is load-bearing.
+   This <div> spent one commit nested inside the `<span class="sub">` of the
+   booth header, because the insertion matched the FIRST `{% if board %}` in
+   the file rather than the block-level one. A <div> inside a <span> is
+   invalid HTML: the parser closes the span implicitly and hoists the div
+   out, orphaning the rest of the sub-line. Three of four cold bug-hunt arms
+   found it and the seat confirmed it in the live document by byte offset.
+   Keep this block between the marks panel and the board form. #}
+{% if is_board %}
+  {# THE BENCH REGISTRY. A bench is a running thing — jackdaw's current bench,
+     talk's current bench, the things that get promoted to Homepage when they
+     are fully deployed. NOT a booth (a booth announces itself and is swept) and
+     NOT a bookmark (a repo page, a model card — those stay on the board below).
+
+     Identity is the NORMALIZED URL, so re-announcing a bench updates its row
+     instead of appending a fifth. `talk` was on the board five times.
+
+     ORDER: state (live → promoted → retired), then name, then id as a total
+     tie-break so two benches sharing a name cannot swap between renders.
+
+     The href is `b.url` — the URL AS POSTED — never `b.id`. The id is
+     normalized for identity; a server that cares about a trailing slash or a
+     case-sensitive path would 404 on it. #}
+  <div class="benches">
+    <div class="bench-head">
+      <span class="bench-title">{{ benches|length }} bench{{ '' if benches|length == 1 else 'es' }}</span>
+      <span class="bench-note">a running thing, registered · re-posting updates the row</span>
+    </div>
+    {% if benches_error %}
+      {# DAMAGED AND ABSENT MUST NOT RENDER THE SAME. Only one of them needs a
+         human, and the v0.2.2 outage was learned by treating them alike. #}
+      <div class="bench-err">the bench registry could not be read: {{ benches_error }}</div>
+    {% elif not benches %}
+      <div class="bench-empty">no benches registered yet — <code>booth bench add &lt;url&gt; &lt;name&gt;</code></div>
+    {% endif %}
+    {% for b in benches %}
+    <div class="bench-row is-{{ b.state }}">
+      <span class="bench-state">{{ b.state }}</span>
+      <div class="bench-main">
+        <a class="bench-link" href="{{ b.url }}" target="_blank" rel="noopener">{{ b.name or b.url }}</a>
+        <div class="bench-url">{{ b.url }}</div>
+      </div>
+      <div class="bench-meta">
+        {% if b.owner %}<span class="bench-who">{{ b.owner }}</span>{% endif %}
+        {# The date it was REGISTERED, not the date it was last touched: `added`
+           survives re-registration and `updated` does not, so `added` is the
+           one that answers "how long has this been around". #}
+        {% if b.added %}<span class="bench-when">{{ b.added[:10] }}</span>{% endif %}
+      </div>
+      <form class="bench-acts" method="post" action="/b/{{ name_url }}/bench-state">
+        <input type="hidden" name="bench" value="{{ b.id }}">
+        {% for s in ("live", "promoted", "retired") %}
+          {% if s != b.state %}
+          <button type="submit" name="state" value="{{ s }}" class="bench-to">{{ s }}</button>
+          {% endif %}
+        {% endfor %}
+        <button type="submit" class="bench-rm" formaction="/b/{{ name_url }}/bench-remove"
+                title="remove this bench">&times;</button>
+      </form>
+    </div>
+    {% endfor %}
+    <form class="bench-add" method="post" action="/b/{{ name_url }}/bench-add">
+      <input type="url" name="url" placeholder="https://host:port/" required>
+      <input type="text" name="name" placeholder="what it is">
+      <button type="submit">register</button>
+    </form>
+  </div>
+{% endif %}
+
 {% if board %}
   {# THE STANDING LINK BOARD. Every agent session on the fleet appends here, so
      this is the one booth where the useful granularity is the ROW, not the
@@ -128,14 +212,17 @@
               formaction="/b/{{ name_url }}/unlink-many">🗑 delete <span id="board-selcount">0</span></button>
     </div>
     {% for e in board %}
-    <div class="board-row{% if e.pinned %} is-pinned{% endif %}">
+    {# DEAD: the row points at a booth that has been swept. 156 of 221 rows.
+       MARKED, never removed — removal is the operator ticking the box and using
+       the bulk control that was already here. #}
+    <div class="board-row{% if e.pinned %} is-pinned{% endif %}{% if e.dead %} board-dead{% endif %}">
       <input class="board-check" type="checkbox" name="sel" value="{{ e.id }}" aria-label="select {{ e.desc }}">
       <button type="submit" class="board-pin{% if e.pinned %} on{% endif %}" formaction="/b/{{ name_url }}/pin"
               name="entry" value="{{ e.id }}" aria-pressed="{{ 'true' if e.pinned else 'false' }}"
               title="{{ 'unpin' if e.pinned else 'pin to top' }}">{{ '★' if e.pinned else '☆' }}</button>
       <div class="board-main">
         <a class="board-link" href="{{ e.url }}" target="_blank" rel="noopener">{{ e.desc }}</a>
-        <div class="board-url">{{ e.url }}</div>
+        <div class="board-url">{{ e.url }}{% if e.dead %} <span class="board-dead-tag">booth is gone</span>{% endif %}</div>
       </div>
       <div class="board-meta">
         {% if e.who %}<span class="board-who">{{ e.who }}</span>{% endif %}
@@ -156,7 +243,40 @@
   {# `elif items` and not a bare `else`: a board booth has NO gallery items (its
      links.md is rendered as the board above and filtered out), so a plain else
      would emit an empty <div class="gallery"> under the board. #}
-  <div class="gallery">
+  {# THE RAIL. Totals and per-filter counts, as LINKS with a query parameter —
+     resolved server-side, so the whole thing works with JavaScript off. The
+     gallery is the surface the operator actually reviews on and U3 already
+     cost the verbatim path its no-JS operation; this one does not repeat that.
+
+     ORDER: the declaration order of FILTERS in app.py. A rail is an ordered
+     collection and invariant 6 binds to it like any other.
+
+     THE GROUP ROW is `rail.groups`, which is EMPTY unless grouping is
+     informative — see `_groups` in app.py. `{% raw %}{% if rail.groups %}{% endraw %}`
+     is therefore the whole guard; the two degenerate cases (one group for
+     everything, one group per item) are decided in Python, where they can be
+     measured, rather than by a count in a template. #}
+  <div class="rail">
+    <span class="rail-total">{{ rail.total }} item{{ '' if rail.total == 1 else 's' }}</span>
+    {% for f in rail.counts %}
+      <a class="rail-f{% if f.key == filter %} on{% endif %}"
+         data-filter="{{ f.key }}"
+         href="/b/{{ name_url }}/{% if f.key != 'all' %}?filter={{ f.key }}{% endif %}"
+         {% if f.key == filter %}aria-current="true"{% endif %}>{{ f.key }} <b>{{ f.n }}</b></a>
+    {% endfor %}
+    {% if rail.groups %}
+    <nav class="rail-groups" aria-label="jump to group">
+      {% for g in rail.groups %}
+        {# The anchor is the first member's EXISTING tile id, so a group has one
+           identity on the page rather than two. Plain fragment links: no JS,
+           and the browser's own back button undoes the jump. #}
+        <a class="rail-g" data-group="{{ g.key }}"
+           href="#{{ g.anchor }}">{{ g.key }} <b>{{ g.n }}</b></a>
+      {% endfor %}
+    </nav>
+    {% endif %}
+  </div>
+  <div class="gallery" id="grid" tabindex="-1">
     {% for it in items %}
     {% if it.doc and it.rendered is not none %}
       {# Docs render INLINE, collapsible, and closable — not a link to a
@@ -187,6 +307,12 @@
           {% else %}
             <pre class="textview doc-body">{{ it.rendered }}</pre>
           {% endif %}
+          {# The doc branch had `markcontrols` and not `marknotes`, so the
+             operator could point at a report and not write down why — on the
+             one item kind whose whole content is prose. Exactly the
+             "patched two of three" failure the blurtoggle macro above was
+             written to prevent, recurring on the macro written to prevent it. #}
+          {{ marknotes(name_url, it, item_marks.get(it.name, [])) }}
         </details>
       </figure>
     {% else %}
@@ -236,6 +362,56 @@
   </div>
 {% endif %}
 
+{% if items %}
+<script id="gridkeys">
+/* GRID KEYBOARD — U7. Additive by construction: every action it reaches is a
+   control that already exists on the tile and already works with a mouse, so
+   the page is complete without this file. It is bound ONLY when there is a
+   grid ({% raw %}{% if items %}{% endraw %} above): binding it on the standing
+   link board would swallow `f` and flag nothing.
+
+   Focus moves in RENDER ORDER, which is the item order filtered by the current
+   filter and never re-sorted — so `→` walks the grid in the same sequence the
+   operator reads it, and the same sequence the zoom ring uses. */
+(function () {
+  var grid = document.getElementById('grid');
+  if (!grid) return;
+  var tiles = function () { return [].slice.call(grid.querySelectorAll('figure.item')); };
+  var at = -1;
+  function focus(i) {
+    var t = tiles();
+    if (!t.length) return;
+    at = Math.max(0, Math.min(i, t.length - 1));
+    t.forEach(function (el, j) { el.classList.toggle('is-cursor', j === at); });
+    t[at].scrollIntoView({ block: 'nearest' });
+  }
+  function current() { var t = tiles(); return at >= 0 && at < t.length ? t[at] : null; }
+  function click(sel) {
+    var el = current(); if (!el) return;
+    var b = el.querySelector(sel); if (b) b.click();
+  }
+  document.addEventListener('keydown', function (e) {
+    /* Never steal a key the operator is typing into a note or a URL bar. */
+    var tag = (e.target.tagName || '').toLowerCase();
+    if (tag === 'input' || tag === 'textarea' || e.target.isContentEditable) return;
+    if (e.metaKey || e.ctrlKey || e.altKey) return;
+    switch (e.key) {
+      case 'ArrowRight': focus(at + 1); e.preventDefault(); break;
+      case 'ArrowLeft':  focus(at <= 0 ? 0 : at - 1); e.preventDefault(); break;
+      case 'f': click('.flagbtn, [name="target"]'); e.preventDefault(); break;
+      case 'n': var el = current();
+                if (el) { var f = el.querySelector('input[type=text], textarea');
+                          if (f) { f.focus(); e.preventDefault(); } }
+                break;
+      case 'Enter': click('a[href^="view"]'); break;
+      case 'Escape':
+        tiles().forEach(function (x) { x.classList.remove('is-cursor'); });
+        at = -1; break;
+    }
+  });
+})();
+</script>
+{% endif %}
 <script>
   /* Copy-to-clipboard for any .copy-btn[data-copy]. The Booth serves over plain
      HTTP on a LAN IP, where navigator.clipboard is undefined (secure-context

booth/templates/doc.html

@@ -33,8 +33,18 @@
     white-space:pre-wrap}
 </style>
 <script>
+  (function () {
+    /* Escape leaves the page, so it must not fire from inside a field someone
+       is typing in — the same guard the image viewer carries, stated in both
+       places because the handler is on `document` in both. */
+    function isEditable(el) {
+      return !!(el && (el.isContentEditable ||
+                       /^(input|textarea|select)$/i.test(el.tagName || '')));
+    }
     document.addEventListener('keydown', function (e) {
+      if (isEditable(e.target)) return;
       if (e.key === 'Escape') window.location.href = {{ ('/b/' ~ name_url ~ '/')|tojson }};
     });
+  })();
 </script>
 {% endblock %}

booth/templates/index.html

@@ -1,4 +1,6 @@
 {% extends "base.html" %}
+{% from "_provenance.html" import provenance %}
+{% from "_lifetime.html" import lifetime %}
 {% block content %}
 <form class="uploader" method="post" action="/upload" enctype="multipart/form-data">
   <label class="drop" for="booth-files">
@@ -38,7 +40,8 @@
       </a>
       <div class="meta">
         <a class="name" href="/b/{{ b.name_url }}/">{{ b.name }}</a>
-        <div class="sub">{{ b.count }} item{{ '' if b.count == 1 else 's' }} · kept · <a class="dl-link" href="/b/{{ b.name_url }}/?download=1" title="download this booth as a zip">⬇ zip</a></div>
+        <div class="sub">{{ b.count }} item{{ '' if b.count == 1 else 's' }} · {{ lifetime(true, b.hold, b.expires_in) }} · <a class="dl-link" href="/b/{{ b.name_url }}/?download=1" title="download this booth as a zip">⬇ zip</a></div>
+        {{ provenance(b.manifest) }}
       </div>
       {# There IS a × here now (operator, 2026-09-21). The old rule was
          release-then-find-it-in-the-other-lane, on the theory that two
@@ -67,11 +70,11 @@
          a label changes width. #}
       <div class="kept-actions">
       <form class="release" method="post" action="/b/{{ b.name_url }}/unkeep"
-            onsubmit="return confirm('Release \u201c{{ b.name }}\u201d?\n\nIt moves to the ephemeral lane so you can wipe it from there. Nothing is deleted by this step.')">
+            data-booth="{{ b.name }}" data-confirm="release">
         <button title="release this board so it can be wiped">release</button>
       </form>
       <form class="wipe wipe-kept" method="post" action="/b/{{ b.name_url }}/delete"
-            onsubmit="return confirm('WIPE the KEPT booth \u201c{{ b.name }}\u201d?\n\nThis deletes it and its files immediately. Kept booths are the ones nothing else will clean up, so nobody else is going to do this for you — and nothing brings it back.')">
+            data-booth="{{ b.name }}" data-confirm="wipe-kept">
         <button title="wipe this KEPT booth now" aria-label="wipe kept booth">×</button>
       </form>
       </div>
@@ -109,7 +112,8 @@
       </a>
       <div class="meta">
         <a class="name" href="/b/{{ b.name_url }}/">{{ b.name }}</a>
-        <div class="sub">{{ b.count }} item{{ '' if b.count == 1 else 's' }} · expires in {{ b.expires_in|dur }} · <a class="dl-link" href="/b/{{ b.name_url }}/?download=1" title="download this booth as a zip">⬇ zip</a></div>
+        <div class="sub">{{ b.count }} item{{ '' if b.count == 1 else 's' }} · {{ lifetime(false, b.hold, b.expires_in) }} · <a class="dl-link" href="/b/{{ b.name_url }}/?download=1" title="download this booth as a zip">⬇ zip</a></div>
+        {{ provenance(b.manifest) }}
       </div>
       {# Promote to the kept lane. The /keep route and the `booth keep` CLI verb
          both predate this button; until 2026-09-19 the UI could only RELEASE a
@@ -120,7 +124,7 @@
         <button title="keep — exempt from the {{ ttl_hours }}h sweep" aria-label="keep booth">★</button>
       </form>
       <form class="wipe" method="post" action="/b/{{ b.name_url }}/delete"
-            onsubmit="return confirm('Wipe booth “{{ b.name }}”?')">
+            data-booth="{{ b.name }}" data-confirm="wipe">
         <button title="wipe now" aria-label="wipe booth">×</button>
       </form>
     </article>
@@ -157,5 +161,39 @@
       }
     });
   })();
+
+  /* Destructive-action confirmation, delegated and DATA-DRIVEN.
+     These were an inline onsubmit calling confirm() with the booth NAME
+     interpolated straight into the JS string literal. Jinja's autoescape is
+     HTML-attribute escaping, not JS-string escaping: the browser decodes the
+     entity back to a quote before the JS parser ever sees it, so a booth name
+     crafted to close that string executed on submit. Booth names are
+     agent-authored — making a folder under the data dir is the whole API — so
+     that is a live path, not a theoretical one.
+
+     The name now travels as a DATA ATTRIBUTE, where escaping is escaping, and
+     never reaches a JS string literal. Same pattern the board controls already
+     use. With JS off the form submits without a prompt, which is what every
+     no-JS browser here already did. */
+  (function () {
+    var WORDS = {
+      release: function (n) {
+        return 'Release \u201c' + n + '\u201d?\n\nIt moves to the ephemeral lane so you '
+             + 'can wipe it from there. Nothing is deleted by this step.';
+      },
+      'wipe-kept': function (n) {
+        return 'WIPE the KEPT booth \u201c' + n + '\u201d?\n\nThis deletes it and its files '
+             + 'immediately. Kept booths are the ones nothing else will clean up, so nobody '
+             + 'else is going to do this for you \u2014 and nothing brings it back.';
+      },
+      wipe: function (n) { return 'Wipe booth \u201c' + n + '\u201d?'; }
+    };
+    document.addEventListener('submit', function (ev) {
+      var form = ev.target.closest ? ev.target.closest('form[data-confirm]') : null;
+      if (!form) return;
+      var word = WORDS[form.getAttribute('data-confirm')];
+      if (word && !confirm(word(form.getAttribute('data-booth') || ''))) ev.preventDefault();
+    }, true);
+  })();
 </script>
 {% endblock %}

booth/templates/marks.html

@@ -1,4 +1,5 @@
 {% extends "base.html" %}
+{% from "_lifetime.html" import lifetime %}
 {% block title %}{{ name }} · marks · The Booth{% endblock %}
 {% block content %}
 {# The marks page for a booth whose own index.html is served VERBATIM. That page
@@ -11,7 +12,7 @@
   {# `marks_open` comes from open_marks() — the ONE openness predicate (INV-2).
      This used to re-derive it in Jinja as `selectattr('answer', 'none')`, which
      read a half-answered pick as closed. #}
-  <span class="sub">{% if marks_open %}<span class="badge badge-mark">{{ marks_open }} open</span> · {% endif %}{{ marks|length }} mark{{ '' if marks|length == 1 else 's' }}</span>
+  <span class="sub">{% if marks_open %}<span class="badge badge-mark">{{ marks_open }} open</span> · {% endif %}{{ marks|length }} mark{{ '' if marks|length == 1 else 's' }} · {{ lifetime(kept, hold, expires_in) }}</span>
 </div>
 {% if marks %}
   {% include "_marks.html" %}

booth/templates/view.html

@@ -99,7 +99,17 @@
     img.addEventListener('load', evaluate);
     window.addEventListener('resize', evaluate);
     if (img.complete) evaluate();
+
+    /* An arrow key inside the note field is a CARET move, not a navigation.
+       The handler is on `document` and the note textarea shipped into this same
+       page, so typing a note and reaching for ← threw the draft away; Escape
+       did it in one keystroke. Anything editable keeps its own keys. */
+    function isEditable(el) {
+      return !!(el && (el.isContentEditable ||
+                       /^(input|textarea|select)$/i.test(el.tagName || '')));
+    }
     document.addEventListener('keydown', function (e) {
+      if (isEditable(e.target)) return;
       if (e.key === 'Escape') window.location.href = BACK;
       else if (e.key === 'ArrowLeft' && PREV) window.location.href = PREV;
       else if (e.key === 'ArrowRight' && NEXT) window.location.href = NEXT;

docs/archive/links-2026-09-22.md

@@ -0,0 +1,234 @@
+# Standing link board — verbatim archive, 2026-09-22
+
+Captured before U6 (benches) shipped, per the ROADMAP rule that a migration
+destroys nothing. 221 rows: 178 booth URLs (156 of them pointing at booths
+already swept) and 43 non-booth rows, 35 distinct after normalization.
+
+U6 itself deletes NOTHING — the dead rows are marked and removal stays the
+operator's two clicks. This archive exists so the board is recoverable
+off-box once he starts pruning, and so the measurements above are checkable
+against the bytes they were taken from.
+
+```markdown
+- [LRPG Authoring Studio — live demo endpoint (ldp-saga)](http://10.100.10.50:8321/Authoring%20Studio.dc.html) <sub>· ldp-dev · 2026-08-19 10:04</sub>
+- [LRPG GM Player — live demo endpoint (ldp-saga; open in iPhone Safari for native)](http://10.100.10.50:8321/GM%20Playback.dc.html) <sub>· ldp-dev · 2026-08-19 10:04</sub>
+- [Scriberr — self-hosted transcription + speaker diarization (ana-ml2 GPU1); also http://10.250.50.54:8080](http://scriberr.ana.internal:8080/) <sub>· infra-ops · 2026-08-23 19:31</sub>
+- [talk — chat with a fleet voice (HTTPS, trusted cert, no warning)](https://talk.nh3.phasefinal.com:8092/) <sub>· tts-dev · 2026-09-06 23:35</sub>
+- [YTVC noise floor A/B — raw vs shipped vs +75 Hz high-pass (2 clips)](http://10.100.10.50:8090/b/ytvc-noise/) <sub>· yt-voice-clipper-dev · 2026-09-09 10:58</sub>
+- [the interview noise floor measured — denoise BEFORE distilling carries 4x better](http://10.100.10.50:8090/b/noise-floor/) <sub>· tts-dev · 2026-09-09 11:00</sub>
+- [the 5 distillation sources staged for professional denoising — drop back as <name>-clean.wav](http://10.100.10.50:8090/b/denoise-in/) <sub>· tts-dev · 2026-09-09 11:01</sub>
+- [hamr: the sliver lever + mutual-block pairing -- the operator four sites at two lever settings (2026-09-09)](http://10.100.10.50:8090/b/hamr-sliver-lever/) <sub>· nh3-dev · 2026-09-09 11:13</sub>
+- [YTVC subtractive denoiser audition — raw vs RNNoise vs DeepFilterNet 3 vs anlmdn, 2 clips + numbers](http://10.100.10.50:8090/b/ytvc-denoise/) <sub>· yt-voice-clipper-dev · 2026-09-09 11:13</sub>
+- [denoise-in — 5 clone sources handed to yt-voice-clipper-dev for a proper deep denoise pass](http://10.100.10.50:8090/b/denoise-in/) <sub>· tts-dev · 2026-09-09 12:49</sub>
+- [hamr: why the gear circle and peak edges read rough -- source vs output vs difference, measured (2026-09-09)](http://10.100.10.50:8090/b/hamr-rough-edges/) <sub>· nh3-dev · 2026-09-09 12:52</sub>
+- [hamr: CLEAN mode rendered on six marks -- and the 1024-vs-4096 test showing my instrument was under-resolved (2026-09-09)](http://10.100.10.50:8090/b/hamr-clean-mode/) <sub>· nh3-dev · 2026-09-09 13:29</sub>
+- [denoise A/B — 5 sources before/after, level-matched; emmie regressed](http://10.100.10.50:8090/b/denoise-ab/) <sub>· tts-dev · 2026-09-09 13:56</sub>
+- [REDO step 1 — pick anchors for lawson/jo/nichols/ana on the cleaned sources (one form)](http://10.100.10.50:8090/b/redo-anchors/asks) <sub>· tts-dev · 2026-09-09 14:01</sub>
+- [hamr: the sliver lever re-rendered at 4x -- the operator width's chunk is real geometry and 7x the default's edge residual (open ask: lever-default)](http://10.100.10.50:8090/b/hamr-sliver-lever/) <sub>· hamr-dev · 2026-09-09 14:31</sub>
+- [hamr: the golden corpus re-rendered at 4x -- the 1024 px instrument inflated edge roughness by 80% on a reading it could not resolve; the colour numbers were never affected](http://10.100.10.50:8090/b/hamr-corpus-4x/) <sub>· hamr-dev · 2026-09-09 14:44</sub>
+- [REDO step 2 — lawson register picks on the cleaned source (7 inline)](http://10.100.10.50:8090/b/redo-lawson/) <sub>· tts-dev · 2026-09-09 14:49</sub>
+- [REDO step 2 — jo register picks on the cleaned source (7 inline)](http://10.100.10.50:8090/b/redo-jo/) <sub>· tts-dev · 2026-09-09 14:49</sub>
+- [REDO step 2 — nichols register picks on the cleaned source (7 inline)](http://10.100.10.50:8090/b/redo-nichols/) <sub>· tts-dev · 2026-09-09 14:49</sub>
+- [REDO step 2 — ana register picks on the cleaned source (7 inline)](http://10.100.10.50:8090/b/redo-ana/) <sub>· tts-dev · 2026-09-09 14:49</sub>
+- [lawson warm rescue — seed axis vs instruction axis (warm is the corpus's untuned string)](http://10.100.10.50:8090/b/lawson-warm/) <sub>· tts-dev · 2026-09-09 15:04</sub>
+- [bank denoise vs source redo — v3+DN hits 49.8 dB; may make the whole redo unnecessary](http://10.100.10.50:8090/b/bank-denoise/) <sub>· tts-dev · 2026-09-09 15:16</sub>
+- [bank denoise A/B — lawson +26 dB, jo +21 dB; 4 of 10 banks would be DAMAGED by it](http://10.100.10.50:8090/b/bank-dn-ab/) <sub>· tts-dev · 2026-09-09 15:31</sub>
+- [Margaery step 1 — anchor picks; ⚠ 15.44s single-clip source, thinnest yet](http://10.100.10.50:8090/b/margaery-anchor/) <sub>· tts-dev · 2026-09-09 15:43</sub>
+- [hamr: 1-2 px regions -- the operator's hue/lightness rule separates 10-25x on his own artwork; lightness does the work; the eye survives at today's default](http://10.100.10.50:8090/b/hamr-thin-regions/) <sub>· hamr-dev · 2026-09-09 15:44</sub>
+- [pewpewstudio web UI restyled on PowerPellet (arcade design system): every screen, dark + daylight (2026-09-09)](http://10.100.10.50:8090/b/pewpew-powerpellet/) <sub>· pewpew-dev · 2026-09-09 15:47</sub>
+- [Margaery — 7 registers x 5 seeds, pick one per register (step 2 of 3)](http://10.100.10.50:8090/b/margaery-registers/) <sub>· tts-dev · 2026-09-09 16:03</sub>
+- [Margaery — denoise A/B on the spliced bank (step 3 of 3)](http://10.100.10.50:8090/b/margaery-denoise/) <sub>· tts-dev · 2026-09-09 16:19</sub>
+- [hamr: the blend-distance gate landed -- the crest's eye ring survives STRIP_WIDTH, the gear's rims and peak's dark-teal strip still go](http://10.100.10.50:8090/b/hamr-thin-regions/) <sub>· hamr-dev · 2026-09-09 17:10</sub>
+- [Breeze — probing the 7 unused direction axes (vendor instructions verbatim)](http://10.100.10.50:8090/b/breeze-axes/) <sub>· tts-dev · 2026-09-09 17:11</sub>
+- [ERP run 7 decision brief — gate failure, exposure, 5 decisions awaiting Vuong](http://10.100.10.50:8090/b/run07-decisions/) <sub>· infra-ops · 2026-09-09 18:10</sub>
+- [R47 tune line runs 4-7 — run 7: length FLAT, RP shape markers moved (quote-first 29%→15%)](http://10.100.10.50:8090/b/r47-runs/) <sub>· brokkr-smithy-dev · 2026-09-09 18:46</sub>
+- [hamr: the blend gate rendered -- the crest's eye ring comes back at STRIP_WIDTH, the gear's rims and peak's strip still go](http://10.100.10.50:8090/b/hamr-thin-regions/) <sub>· hamr-dev · 2026-09-09 18:52</sub>
+- [Tag sweep redone — leak test = vocabulary test; the ear questions](http://10.100.10.50:8090/b/tag-sweep/) <sub>· tts-dev · 2026-09-09 22:49</sub>
+- [hamr henge/66 peak: which site is 'the chunk' -- ask + the four candidate sites](http://10.100.10.50:8090/b/hamr-henge66-peak/) <sub>· hamr-dev · 2026-09-09 23:13</sub>
+- [Chunk seams A/B — paragraph-only chunking, and the render ceiling is lower than we thought](http://10.100.10.50:8090/b/chunk-seams/) <sub>· tts-dev · 2026-09-09 23:29</sub>
+- [hamr peak: the operator's chunk (the small peak's left face) -- under the size levers, before/after the apex unit](http://10.100.10.50:8090/b/hamr-peak-left-face/) <sub>· hamr-dev · 2026-09-10 07:26</sub>
+- [hamr: the peak's halo -- the tint reach null, the sliver lever, the support rule (henge/66 third rule)](http://10.100.10.50:8090/b/hamr-peak-halo/) <sub>· hamr-dev · 2026-09-10 07:58</sub>
+- [Level decay is LENGTH-driven, not soft/whisper — every direction collapses at 1400 chars](http://10.100.10.50:8090/b/level-decay/) <sub>· tts-dev · 2026-09-10 08:47</sub>
+- [BabyBronte voice A/B — base vs H02 LoRA on 9 neutral prompts, 2 seeds each](http://10.100.10.50:8090/b/babybronte-voice/) <sub>· infra-ops · 2026-09-10 15:08</sub>
+- [hamr: the 2.5 fold -- frame closing fix, the O(N) vote (byte-identical peak), the VMDE engine document read against hamr](http://10.100.10.50:8090/b/hamr-2-5-fold/) <sub>· hamr-dev · 2026-09-10 15:57</sub>
+- [hamr: the region-energy segmenter spike (henge 71) -- the Potts prior in the vote's seat, against the landed 2.5](http://10.100.10.50:8090/b/hamr-region-energy/) <sub>· hamr-dev · 2026-09-10 16:05</sub>
+- [BabyBronte rung 2 — 1.7B base vs 1.7B tuned vs 0.6B tuned, 9 prompts, 2 seeds](http://10.100.10.50:8090/b/babybronte-1p7b/) <sub>· infra-ops · 2026-09-10 22:38</sub>
+- [hamr: the state of the pipeline at c18c4e1 (v1.3.0 + the hygiene unit) -- seven reference marks and the synthetic corpus, source | 1x | 4x](http://10.100.10.50:8090/b/hamr-state-2026-09-11/) <sub>· hamr-dev · 2026-09-10 23:20</sub>
+- [BabyBronte rung 3 — 4B base vs 4B tuned vs 1.7B tuned, + the Abernathy frame prompt](http://10.100.10.50:8090/b/babybronte-4b/) <sub>· infra-ops · 2026-09-11 05:35</sub>
+- [bragi :8196 — the fleet direction layer, LIVE 2026-09-11 (U1 null director, +2.32ms TTFA cost, cap 6400)](http://irv-ml1.nh3.internal:8196/health) <sub>· nh3-dev · 2026-09-11 05:47</sub>
+- [BabyBronte rung 3 (step-75 recut) — 4B base vs 4B tuned vs 1.7B, + frame and embedded-instruction prompts](http://10.100.10.50:8090/b/babybronte-4b/) <sub>· infra-ops · 2026-09-11 05:54</sub>
+- [Bragi U2 spike — blinded 5-arm fast-director audition, 7 inline asks, ear verdict gates U2](http://10.100.10.50:8090/b/bragi-u2-spike/) <sub>· nh3-dev · 2026-09-11 06:00</sub>
+- [Skaldsong beat→paragraph — 10 formats on the adapted 4B vs an instruct model, + stitched story](http://10.100.10.50:8090/b/skaldsong-beats/) <sub>· infra-ops · 2026-09-11 06:24</sub>
+- [hamr state booth at a7ee4ab: seven reference marks + sixteen synthetic cases, source | 1x | 4x, after the ridge-order and test-hygiene units](http://10.100.10.50:8090/b/hamr-state-2026-09-11-a7ee4ab/) <sub>· hamr-dev · 2026-09-11 08:41</sub>
+- [hamr state booth, clean mode default (colour_geometry 3.13): only the crest's white tick changes against a7ee4ab](http://10.100.10.50:8090/b/hamr-state-2026-09-11-clean/) <sub>· hamr-dev · 2026-09-11 10:23</sub>
+- [hamr run_smoothing 2.2, the corner core: circuit/gear/peak/crest/vastblue at the new corner rule, with corner overlays](http://10.100.10.50:8090/b/hamr-corner-core/) <sub>· hamr-dev · 2026-09-11 11:12</sub>
+- [hamr regularizer 3.0, the run solve (U7 on runs): circuit/gear/peak/crest/vastblue after the stretch pool and solve, with the circuit site the first form broke](http://10.100.10.50:8090/b/hamr-run-solve/) <sub>· hamr-dev · 2026-09-11 13:50</sub>
+- [hamr regularizer 3.1, the junction at the meet: the circuit's pads 3.0 vs 3.1 and the five marks](http://10.100.10.50:8090/b/hamr-run-solve-31/) <sub>· hamr-dev · 2026-09-11 15:02</sub>
+- [BabyYarros eval — voice A/B + beat→paragraph + delta_cb (Base@125 vs Instruct vs base control)](http://10.100.10.50:8090/b/babyyarros-voice/) <sub>· infra-ops · 2026-09-11 15:59</sub>
+- [bifrost 1.2.0 on the gitea PyPI index — wire v0.8 memory.* record profile (#17)](https://gitea.phasefinal.com/vh/-/packages/pypi/bifrost/1.2.0) <sub>· bifrost-dev · 2026-09-11 16:58</sub>
+- [bifrost #17 — wire v0.8 record profile (adoption arc, gates, release)](https://gitea.phasefinal.com/vh/bifrost/issues/17) <sub>· bifrost-dev · 2026-09-11 16:58</sub>
+- [bifrost 1.2.1 — supplement-fold patch (explicit record-engine guards; descriptor ownership boundary)](https://gitea.phasefinal.com/vh/-/packages/pypi/bifrost/1.2.1) <sub>· bifrost-dev · 2026-09-11 17:32</sub>
+- [BabyYarros — Janis beat: 4 prompt arms x 4 seeds, beat->paragraph formula fitting](http://10.100.10.50:8090/b/babyyarros-janis/) <sub>· infra-ops · 2026-09-11 21:15</sub>
+- [hamr on five fresh arbo marks (owl, bee, rocket, wolf, lantern) -- landed pipeline, clean mode, 1x + 4x](http://10.100.10.50:8090/b/hamr-arbo-logos/) <sub>· hamr-dev · 2026-09-11 22:35</sub>
+- [FV colo on-site playbook — print before the trip (OPNsense + fv-ml1, anti-lockout)](http://10.100.10.50:8090/b/fv-onsite/) <sub>· infra-ops · 2026-09-12 07:54</sub>
+- [hamr arbo marks AFTER colour_decomposition 2.10 (the interior-ends tint reading): owl before/after, the four others byte-identical](http://10.100.10.50:8090/b/hamr-arbo-logos-2/) <sub>· hamr-dev · 2026-09-12 07:55</sub>
+- [hamr: the midline rule (colour_geometry 3.14) on the owl -- source | before | midline | far, 4x, and the runs the instrument flagged](http://10.100.10.50:8090/b/hamr-midline/) <sub>· hamr-dev · 2026-09-12 22:18</sub>
+- [Qwen3.8-Flash-Next ABLITERATED NVFP4 + FP8 PLE — candidate for the fv-ml1 single-card gen seat](https://huggingface.co/dealignai/Qwen3.8-Flash-Next-ABLITERATED-NVFP4) <sub>· infra-ops · 2026-09-12 22:21</sub>
+- [vLLM canonical Qwen3.8-Flash-Next recipe — PLE CPU-offload + the don't-enable-MTP measurement](https://recipes.vllm.ai/Qwen/Qwen3.8-Flash-Next/) <sub>· infra-ops · 2026-09-12 22:21</sub>
+- [hamr: FAR shipped (colour_geometry 3.16) -- the five arbo marks before | after at 4x, and the per-run instrument](http://10.100.10.50:8090/b/hamr-far/) <sub>· hamr-dev · 2026-09-13 00:13</sub>
+- [hamr: edge-pixel rule spike -- census overlays (third-layer boundary pixels, green explained / red not) and the geometry arms](http://10.100.10.50:8090/b/hamr-edge-pixels/) <sub>· hamr-dev · 2026-09-13 09:28</sub>
+- [hamr: colour_geometry 3.17 the line clause -- crest eye ring gone, lens kept; owl / circuit / lantern byte-identical at 4x](http://10.100.10.50:8090/b/hamr-width-clause/) <sub>· hamr-dev · 2026-09-13 14:00</sub>
+- [hamr: the golden corpus at colour_geometry 3.17 (the line clause) -- seven reference marks, faces and runs, source | 1x | 4x](http://10.100.10.50:8090/b/hamr-corpus-3.17/) <sub>· hamr-dev · 2026-09-13 16:53</sub>
+- [hamr: the DXF cut document beside the SVG runs profile on the seven corpus marks (source | SVG | DXF, 1x and 4x zooms; .dxf files alongside)](http://10.100.10.50:8090/b/hamr-dxf/) <sub>· hamr-dev · 2026-09-13 23:18</sub>
+- [Flash-Next gen-large candidate #1: abliterated + W4A16 weight-only experts + FP8 PLE; blocked only by a missing ple_embedding_dtype config key](https://huggingface.co/gorbatjovy/qwen3.8-flash-next-abliterated-NVFP4-plefp8) <sub>· infra-ops · 2026-09-14 02:16</sub>
+- [Flash-Next gen-large candidate #2: fully weight-only (W4A16 experts + FP8_PB_WO dense), loads as-is, but NOT abliterated](https://huggingface.co/lovedheart/Qwen3.8-Flash-Next-NVFP4-W4A16-4-Over-6-FP8) <sub>· infra-ops · 2026-09-14 02:16</sub>
+- [cyberprev-27b — abliterated Qwen3.8-27B sec seat (fv-ml1 GPU0, dflash k=7), replaced sentinel-r3](http://10.251.50.54:8025/docs) <sub>· infra-ops · 2026-09-14 04:29</sub>
+- [hamr-server 1.5: the SPA booth pass with Download DXF (state 08b) and the refused-selection state re-pinned to server 1.6](http://10.100.10.50:8090/b/hamr-server-1.5/) <sub>· hamr-dev · 2026-09-14 10:22</sub>
+- [hamr web front end UI brief (requirements and flow for a design system; also docs/design/ui-brief.md)](https://claude.ai/code/artifact/eae98fde-784b-4f4d-b0e3-c87a229da564) <sub>· hamr-dev · 2026-09-14 10:25</sub>
+- [https://claude.ai/code/artifact/eae98fde-784b-4f4d-b0e3-c87a229da564](https://claude.ai/code/artifact/eae98fde-784b-4f4d-b0e3-c87a229da564) <sub>· hamr-dev · 2026-09-14 10:25</sub>
+- [hamr web front end UI brief, boothed (kept): index.html + ui-brief.md](http://10.100.10.50:8090/b/hamr-ui-brief/) <sub>· hamr-dev · 2026-09-14 10:56</sub>
+- [pewpewstudio web front end UI brief, boothed (kept): index.html + ui-brief.md + the integration package (tarball + fixtures)](http://10.100.10.50:8090/b/pewpew-ui-brief/) <sub>· pewpew-dev · 2026-09-14 12:42</sub>
+- [pewpewstudio web front end UI brief (flow, shape, requirements for a design agent; also docs/design/ui-brief.md)](https://claude.ai/code/artifact/281bcdc7-bcce-46d7-b0ca-ec90df22151f) <sub>· pewpew-dev · 2026-09-14 12:42</sub>
+- [Headscale: Tailscale setup for macOS/iOS/tvOS — GUI steps + downloadable config profiles](https://headscale.phasefinal.com/apple) <sub>· infra-ops · 2026-09-14 13:49</sub>
+- [pewpewstudio: the UI blueprint vendored (Claude Design handoff from booth 28-indigo) -- provenance, state inventory, fidelity notes; source at docs/design/blueprint/](http://10.100.10.50:8090/b/pewpew-ui-brief/blueprint/README.md) <sub>· pewpew-dev · 2026-09-14 18:38</sub>
+- [pewpewstudio web: the blueprint implemented -- one still per surface per state (67), cabinet + daylight](http://10.100.10.50:8090/b/pewpew-blueprint/) <sub>· pewpew-dev · 2026-09-14 20:55</sub>
+- [hamr: the C kernel for the cubic fit -- where its geometry differs from 2.4 (4x panels) and the ask on the gate](http://10.100.10.50:8090/b/hamr-cubic-kernel/) <sub>· hamr-dev · 2026-09-14 22:28</sub>
+- [Homepage — Parakeet ASR card now live under AI - Audio Tools (fv-ml1 GPU 3, :8300)](http://10.0.50.45:5100/) <sub>· nh3-dev · 2026-09-15 01:41</sub>
+- [talk v10 — Sindra with ears: push-to-talk STT via ext-stt + barge-in (nh3-dev)](https://talk.nh3.phasefinal.com:8092/) <sub>· nh3-dev · 2026-09-15 08:27</sub>
+- [talk v10 — the fleet speaks AND listens (Grima push-to-talk + barge-in)](https://talk.nh3.phasefinal.com:8092/) <sub>· nh3-dev · 2026-09-15 08:28</sub>
+- [Open-weight releases landscape scan 2026-09-15 — LLM/image/TTS, ranked + licenses verified](https://gitea.phasefinal.com/vh/brokkr-smithy/src/commit/6adcde6/research/landscape-scans/open-weight-releases-2026-09-15.md) <sub>· brokkr-scan-dev · 2026-09-15 09:20</sub>
+- [ldp-saga — voice-over step with authored words: GM stage (iPhone) + Studio drawer screenshots](http://10.100.10.50:8090/b/ldp-vo-body/) <sub>· ldp-dev · 2026-09-15 11:31</sub>
+- [talk PREVIEW (v11 unreleased) — kiosk persona + prompt library + hands-free VAD; http so no mic](http://10.100.10.50:8095/) <sub>· nh3-dev · 2026-09-15 14:08</sub>
+- [talk v12 LIVE — hands-free VAD + 4 personas (assistant/sindra/narrator/kiosk) + Grima STT](https://talk.nh3.phasefinal.com:8092/) <sub>· nh3-dev · 2026-09-15 14:13</sub>
+- [talk v12 — internal IP (accept the cert warning; wildcard covers names, not IPs). Hands-free + 4 personas.](https://10.100.10.50:8092/) <sub>· nh3-dev · 2026-09-15 14:18</sub>
+- [hamr circuit: census of thin surviving regions, source|1x|4x per site (2026-09-16)](http://10.100.10.50:8090/b/hamr-circuit-slivers/) <sub>· hamr-dev · 2026-09-15 15:03</sub>
+- [hamr circuit: the full cut file (SVG runs profile + DXF) on white, 1x and 4x whole (2026-09-16)](http://10.100.10.50:8090/b/hamr-dxf/) <sub>· hamr-dev · 2026-09-15 15:10</sub>
+- [hamr owl (arbo 00-seed7777): the full cut file on white, 1x and 4x (2026-09-16)](http://10.100.10.50:8090/b/hamr-owl-cut/) <sub>· hamr-dev · 2026-09-15 15:17</sub>
+- [hamr circuit: the ten arrowed sites (possum-51), source | faces 4x | runs 4x, with the runs and junctions at each (2026-09-16)](http://10.100.10.50:8090/b/hamr-circuit-arrows/) <sub>· hamr-dev · 2026-09-15 15:17</sub>
+- [hamr: the owl before/after the shade rule (colour_decomposition 2.12), the four arrowed sites at 1x and 4x](http://10.100.10.50:8090/b/hamr-owl-shades/) <sub>· hamr-dev · 2026-09-15 20:01</sub>
+- [hamr unit 2: the circuit's edge teeth before/after (colour_geometry 3.27) -- the ten arrowed sites and two interior seam sites, SOURCE | before | after at 1x and 4x](http://10.100.10.50:8090/b/hamr-circuit-teeth/) <sub>· hamr-dev · 2026-09-15 22:57</sub>
+- [hamr 3.27: every thin excursion the clause reads on twenty marks at the pixel bar (175 panels; GOES/stays in each caption)](http://10.100.10.50:8090/b/hamr-excursions-f10/) <sub>· hamr-dev · 2026-09-15 22:57</sub>
+- [BabyYarros beat→paragraph: same beat, 4 arms (base / raw-text / pair-SFT 2ep / 3ep)](http://10.100.10.50:8090/b/babyyarros-beats/) <sub>· infra-ops · 2026-09-16 07:29</sub>
+- [hamr v2 S0: the smoother's chain vs potrace's fallback on every refused mono node of the eight marks, worst site per node at 4x (2026-09-16)](http://10.100.10.50:8090/b/hamr-v2-s0-smoother/) <sub>· hamr-dev · 2026-09-16 08:55</sub>
+- [hamr U0 — the truth-corpus acceptance gate: 24 conditions, potrace 3x vs the extractor's iso-contours, table + overlays at 1x and 4x](http://10.100.10.50:8090/b/hamr-u0-acceptance/) <sub>· hamr-dev · 2026-09-16 11:11</sub>
+- [hamr acceptance 1.2 verdict table -- 24 conditions, three arms over the raster per condition (from hamr-dev's fold of two Heid panels)](http://10.100.10.50:8090/b/hamr-u0-acceptance/) <sub>· heid · 2026-09-16 11:17</sub>
+- [Assistant voice — accent calibration: 7 endpoints from the existing battery, inline ask](http://10.100.10.50:8090/b/assistant-accent/) <sub>· nh3-dev · 2026-09-16 11:18</sub>
+- [Assistant voice — the blend n=5, matched-seed triples vs both endpoints](http://10.100.10.50:8090/b/assistant-blend/) <sub>· nh3-dev · 2026-09-16 11:20</sub>
+- [Peedlar repo (photo → eBay/FB Marketplace listing metadata) — minted 2026-09-16](https://gitea.phasefinal.com/vh/peedlar) <sub>· nh3-dev · 2026-09-16 11:26</sub>
+- [Sun and Sea Pro — concept tiles A/B/C + the rulings ask (design-systems)](http://10.100.10.50:8090/b/sunsea/) <sub>· design-dev · 2026-09-16 11:35</sub>
+- [Peedlar — UI design brief + northstar/frame/invariants/interview record (vor-ui pass 2026-09-16)](http://10.100.10.50:8090/b/peedlar-design-brief/) <sub>· peedlar-dev · 2026-09-16 14:01</sub>
+- [hamr U1 the tracer skeleton (tracer 3.0): the v2 tree over the eight marks with ids and holes, potrace beside it, 4x windows, the rule fixtures](http://10.100.10.50:8090/b/hamr-u1-tracer/) <sub>· hamr-dev · 2026-09-16 14:23</sub>
+- [Peedlar — vor-plan draft bundle (plan, frame, invariants, northstar, record) for teardown, 2026-09-16](http://10.100.10.50:8090/b/peedlar-plan-draft/) <sub>· peedlar-dev · 2026-09-16 16:17</sub>
+- [Peedlar — spike R-4 report: gen schema adherence, 180/180 valid (2026-09-16)](http://10.100.10.50:8090/b/peedlar-spike-r4/) <sub>· peedlar-dev · 2026-09-16 17:18</sub>
+- [hamr U2 (ir 7.0): the mono SVG before/after the IR moved onto points, eight marks, 1x and 4x](http://10.100.10.50:8090/b/hamr-u2-ir/) <sub>· hamr-dev · 2026-09-16 17:22</sub>
+- [JackDAW audition bench — live HEAD of main (self-signed HTTPS, one-time trust prompt)](https://10.100.10.50:4500/) <sub>· jackdaw-dev · 2026-09-16 18:32</sub>
+- [Peedlar UI in Sun and Sea Pro — nine surfaces + DESIGN.md (design-systems, for peedlar-dev)](http://10.100.10.50:8090/b/peedlar-ui/) <sub>· design-dev · 2026-09-16 19:27</sub>
+- [Assistant anchor — rp-s113 vs the existing emily, collision check before building a bank](http://10.100.10.50:8090/b/assistant-anchor/) <sub>· nh3-dev · 2026-09-16 19:29</sub>
+- [imogen — register bank ear gate before freezing (5 registers off rp-s113)](http://10.100.10.50:8090/b/imogen/) <sub>· nh3-dev · 2026-09-16 19:45</sub>
+- [imogen — gentle + dry re-roll, 3 draws each vs the rejected originals](http://10.100.10.50:8090/b/imogen-reroll/) <sub>· nh3-dev · 2026-09-16 19:50</sub>
+- [Peedlar — spike R-3 report: split heuristic on the cedarwood-4 pile (pairwise VLM + identify-and-merge, 4-image cap), 2026-09-16](http://10.100.10.50:8090/b/peedlar-spike-r3/) <sub>· peedlar-dev · 2026-09-16 19:53</sub>
+- [hamr U4: the colour spine on owner fields at 1x -- v1.6.1 (3x potrace) vs colour_spine 3.0, eight marks, 1x + 4x diff windows, the 1x/3x A/B table](http://10.100.10.50:8090/b/hamr-u4-readers/) <sub>· hamr-dev · 2026-09-16 20:26</sub>
+- [imogen LIVE — voice 22 on the roster, all five registers through the gateway](http://10.100.10.50:8090/b/imogen-live/) <sub>· nh3-dev · 2026-09-16 20:34</sub>
+- [talk v15 — imogen is the default voice; 22 voices, 4 personas, hands-free](https://talk.nh3.phasefinal.com:8092/) <sub>· nh3-dev · 2026-09-16 20:39</sub>
+- [Peedlar v0.1.0 — U0 scaffold deployed on nh3-dev (health placeholder SPA + /healthz)](http://10.100.10.50:8094/) <sub>· peedlar-dev · 2026-09-16 23:42</sub>
+- [hamr U3: the mono smoothing -- every refused node's chain (blue) beside the polyline it replaces (red), eight marks, 1x and 4x](http://10.100.10.50:8090/b/hamr-u3-mono-smoothing/) <sub>· hamr-dev · 2026-09-17 00:11</sub>
+- [2026-09-17 Civitai batch A/B — 6 promotion/retirement decisions, inline asks (comfy-dev)](http://10.100.10.50:8090/b/civitai-20260917-ab/) <sub>· comfy-dev · 2026-09-17 01:49</sub>
+- [Breeze v5 vendor-pin rebase — A/B clips, gate numbers, two decisions](http://10.100.10.50:8090/b/breeze-v5-gate/) <sub>· tts-dev · 2026-09-17 02:36</sub>
+- [ldp-saga U4 — control panel + bootstrap view screenshots (polish-pass input)](http://10.100.10.50:8090/b/ldp-u4-panel/) <sub>· ldp-dev · 2026-09-17 02:38</sub>
+- [lv voices four arms — same beat, same neutral prompt: control vs Bronte vs Yarros vs Hemingway (2026-09-17)](http://10.100.10.50:8090/b/lv-voices-four-arms/) <sub>· infra-ops · 2026-09-17 07:52</sub>
+- [hamr U6: the eight marks' faces and cut on white, v1.6.1 (potrace) beside main (own tracer), 1x + 4x worst window, trace timings](http://10.100.10.50:8090/b/hamr-u6-before-after/) <sub>· hamr-dev · 2026-09-17 08:06</sub>
+- [ldp-demo-kit 2026-09-17-0816 (build 99040b2): VO authored words in Eric's kit](http://10.100.10.50:8090/b/ldp-demo-kit/) <sub>· ldp-dev · 2026-09-17 08:17</sub>
+- [hamr U6 regression sites: crest/circuit/owl difference clusters at 4x, SOURCE | v1.6.1 | main | candidate (coverage-field evidence)](http://10.100.10.50:8090/b/hamr-u6-sites/) <sub>· hamr-dev · 2026-09-17 08:33</sub>
+- [talk favicon commission — comfy-dev raster candidates, hamr-dev SVG trace](http://10.100.10.50:8090/b/talk-favicon/) <sub>· tts-dev · 2026-09-17 08:43</sub>
+- [Peedlar U2 ingest screen — four phone states from a real headless Chromium run](http://10.100.10.50:8090/b/peedlar-u2/) <sub>· nh3-dev · 2026-09-17 09:19</sub>
+- [Peedlar v0.2.3 live — U2 ingest: photograph a pile from a phone, send it, top an item up](http://10.100.10.50:8094/) <sub>· nh3-dev · 2026-09-17 10:15</sub>
+- [Peedlar v0.2.4 live — U2 ingest, all three review rounds folded (17 defects)](http://10.100.10.50:8094/) <sub>· nh3-dev · 2026-09-17 11:04</sub>
+- [hamr circuit: the five sites where main's runs depart from v1.6.1's (SOURCE | v1 | main at 4x)](http://10.100.10.50:8090/b/hamr-u6-departures/) <sub>· hamr-dev · 2026-09-17 11:05</sub>
+- [hamr circuit: the trace-to-pad corners on both trees at 4x -- the indented-lines family](http://10.100.10.50:8090/b/hamr-u6-dents/) <sub>· hamr-dev · 2026-09-17 11:05</sub>
+- [Peedlar ingest UI — before/after in six states, with an open ask on fonts + pricing pills](http://10.100.10.50:8090/b/peedlar-ui-polish/) <sub>· design-dev · 2026-09-17 11:25</sub>
+- [hamr run_smoothing 3.4: the chord-of-a-curve clause -- the circuit's pads and trace ends as lines, before/after at 6x](http://10.100.10.50:8090/b/hamr-short-stretches/) <sub>· hamr-dev · 2026-09-17 11:56</sub>
+- [ldp-demo-kit 2026-09-17-1243 (a912928): Eric's 09-17 canonical + VO words — install this one](http://10.100.10.50:8090/b/ldp-demo-kit/) <sub>· ldp-dev · 2026-09-17 12:43</sub>
+- [Peedlar v0.2.5 — surface 1 dressed in Sun and Sea Pro (design-dev), four phone states](http://10.100.10.50:8090/b/peedlar-u2-design/) <sub>· nh3-dev · 2026-09-17 15:54</sub>
+- [talk favicon — the traced mark (B) and its 16/32/64px proof](http://10.100.10.50:8090/b/talk-favicon/) <sub>· nh3-dev · 2026-09-17 15:58</sub>
+- [Peedlar v0.3.0 — the first release a seller can use (ingest + top-up; split is U3)](https://gitea.phasefinal.com/vh/peedlar/releases/tag/v0.3.0) <sub>· nh3-dev · 2026-09-17 15:59</sub>
+- [hamr corner response A/B: 3.4 as landed vs the capped response by angle -- the circuit's bends, the crest's and gear's small fillets](http://10.100.10.50:8090/b/hamr-corner-ab/) <sub>· hamr-dev · 2026-09-17 17:24</sub>
+- [ldp-demo-kit 2026-09-17-1752 (a28e8d5): Eric's 09-17 canon + VO words + GM Markdown subset](http://10.100.10.50:8090/b/ldp-demo-kit/ldp-demo-kit-2026-09-17-1752.zip) <sub>· ldp-dev · 2026-09-17 17:52</sub>
+- [Sun and Sea Pro v1.1.0 — rulings + the Peedlar ingest before/after that started it](http://10.100.10.50:8090/b/peedlar-ui-polish/) <sub>· design-dev · 2026-09-17 17:59</sub>
+- [ldp-demo-kit 2026-09-17-1804 (265a3ad): + _underline_](http://10.100.10.50:8090/b/ldp-demo-kit/ldp-demo-kit-2026-09-17-1804.zip) <sub>· ldp-dev · 2026-09-17 18:04</sub>
+- [ldp-saga — GM Markdown subset samples (source + renders)](http://10.100.10.50:8090/b/ldp-markdown/) <sub>· ldp-dev · 2026-09-17 18:06</sub>
+- [hamr colour_spine 3.7, the paired witness: circuit arrows 1-3 at 12x, every departure site before/after at 1x+4x, the crest's eye](http://10.100.10.50:8090/b/hamr-witness/) <sub>· hamr-dev · 2026-09-17 18:48</sub>
+- [Dragonfire Acoustics — three concept directions + the five rulings that gate the build](http://10.100.10.50:8090/b/dfa-concepts/) <sub>· design-dev · 2026-09-17 18:49</sub>
+- [Dragonfire Acoustics — sample landing page, standalone HTML for client screenshots](http://10.100.10.50:8090/b/dfa-landing/) <sub>· design-dev · 2026-09-17 18:58</sub>
+- [hamr run_smoothing 3.5, the corner response by angle between two stretches: circuit arrows 2-3 and new corners, crest's curves unkinked, at 8x](http://10.100.10.50:8090/b/hamr-corner-guard/) <sub>· hamr-dev · 2026-09-17 18:59</sub>
+- [hamr: golden corpus on main 53356c5, faces and cut on white, 1x sheets and 4x wholes](http://10.100.10.50:8090/b/hamr-corpus-2026-09-18/) <sub>· hamr-dev · 2026-09-17 21:51</sub>
+- [Peedlar surface 2 — a live split of the R-3 pile, ready to confirm (U3)](http://10.100.10.50:8094/batches/6fb2952b-e3b1-4fbd-9694-5f3f3f5d75d0/split) <sub>· nh3-dev · 2026-09-18 07:08</sub>
+- [Peedlar surface 2 — a scratch split to poke at (merge/split/move/drop/restore all live)](http://10.100.10.50:8094/batches/a7058924-a855-40b3-bfc5-11f3f258df27/split) <sub>· nh3-dev · 2026-09-18 07:13</sub>
+- [Peedlar U3 — surface 2 on desk and phone, plus an interaction run](http://10.100.10.50:8090/b/peedlar-u3/) <sub>· nh3-dev · 2026-09-18 07:16</sub>
+- [tag placement A/B — does moving (giggle) stop it overlapping the next line? (ask inside)](http://10.100.10.50:8090/b/tag-placement/) <sub>· tts-dev · 2026-09-18 07:17</sub>
+- [seam gap audition — 0-500ms between generations, 11 arms (ask inside)](http://10.100.10.50:8090/b/seam-gap/) <sub>· tts-dev · 2026-09-18 07:27</sub>
+- [FleetTools index lives at ~/FLEETTOOLS.md on nh3-dev — agent-family-agnostic fleet capability map](http://10.100.10.50:8090/) <sub>· nh3-dev · 2026-09-18 07:35</sub>
+- [Peedlar v0.4.0 — the split ships; capability 1 of five is MET](http://10.100.10.50:8094/) <sub>· nh3-dev · 2026-09-18 08:54</sub>
+- [talk favicon — inverted, transparent, before/after proof at 4 sizes](http://10.100.10.50:8090/b/talk-favicon/) <sub>· tts-dev · 2026-09-18 13:53</sub>
+- [ShutterChute macOS app icon — 3 variants + the 16px proof sheets (comfy-dev, for shutter-dev)](http://10.100.10.50:8090/b/shutterchute-icon/) <sub>· comfy-dev · 2026-09-18 14:00</sub>
+- [NH3↔Anaheim mesh now DIRECT (was DERP-relayed): cross-site HTTP 1.2s→0.015s, STT 1.4s→0.25s — ana-gw UDP 41641 port-forward 2026-09-18](http://10.100.10.50:8090/b/links/) <sub>· nh3-dev · 2026-09-18 14:17</sub>
+- [talk favicon — three-way blue comparison (live vs page accent vs comfy remake)](http://10.100.10.50:8090/b/talk-favicon/) <sub>· tts-dev · 2026-09-18 14:26</sub>
+- [DNS fixed fleet-wide 2026-09-18: cross-site resolver ring + AdGuard ratelimit 20-per-/24 set to 0 — .internal stalls 1-in-8 to zero](http://10.100.10.50:8090/b/links/) <sub>· nh3-dev · 2026-09-18 14:35</sub>
+- [ShutterChute on Paula's mini (v0.9.7) — session token rotates on every restart, read it from /Users/Shared/shutterchute/app.url or the deploy output](http://10.100.10.50:8477/) <sub>· shutter-dev · 2026-09-18 14:46</sub>
+- [asking arbo vs directing it — both icon commissions re-run on the corrected chain, with the 16px verdicts](http://10.100.10.50:8090/b/arbo-asked/) <sub>· comfy-dev · 2026-09-18 14:54</sub>
+- [Blind A/B/C: is Imogen's 39.96s register bank worth 116ms a turn? (breeze v8)](http://10.100.10.50:8090/b/imogen-register/) <sub>· tts-dev · 2026-09-18 20:30</sub>
+- [Sindra identity scouting — 5 SFW/NSFW pairs on moody-krea2 (comfy-dev, for adhoc-agent)](http://10.100.10.50:8090/b/sindra-face-1/) <sub>· comfy-dev · 2026-09-19 12:44</sub>
+- [Sindra casting — 5 different women, 2 fixed scenes (gym / beach), comfy-dev](http://10.100.10.50:8090/b/sindra-cast/) <sub>· comfy-dev · 2026-09-19 15:25</sub>
+- [the three MiniMax Music 3 songs (Aug 2026) — recovered from render scratch, kept, captions carry the recovered lyrics](http://10.100.10.50:8090/b/music3-songs/) <sub>· comfy-dev · 2026-09-19 15:26</sub>
+- [Sindra A — curvier stepped across 4 levels, face frozen (comfy-dev)](http://10.100.10.50:8090/b/sindra-curve/) <sub>· comfy-dev · 2026-09-19 15:36</sub>
+- [the settled Sindra — 5 SFW environments + 5 NSFW poses, identity block verbatim (comfy-dev)](http://10.100.10.50:8090/b/sindra-set/) <sub>· comfy-dev · 2026-09-19 15:44</sub>
+- [NVV markers by ear: is (chuckle) real? + the leak test is dead on breeze v8](http://10.100.10.50:8090/b/nvv-probe/) <sub>· tts-dev · 2026-09-19 17:05</sub>
+- [tts-bench — type/direct/render against the live TTS seat (voice picker, custom directions, marker palette)](http://nh3-dev.nh3.internal:8095/) <sub>· tts-dev · 2026-09-19 17:14</sub>
+- [Sindra voice audition (adhoc-agent commission) — designed synthetic, 3 registers x 2 takes + polyglot probe](http://10.100.10.50:8090/b/sindra-voice-1/) <sub>· tts-dev · 2026-09-19 22:49</sub>
+- [Sindra ANCHOR field — n=15 on the intimate prompt, 13 in the 8-10s window, pick one to freeze](http://10.100.10.50:8090/b/sindra-anchor/) <sub>· tts-dev · 2026-09-19 22:55</sub>
+- [Sindra is LIVE — new designed voice replaces the NZ contralto; bank vs anchor A/B inside](http://10.100.10.50:8090/b/sindra-live/) <sub>· tts-dev · 2026-09-19 23:14</sub>
+- [Cicada repo (was Imogen) — embodied voice assistant, design bundle + embodiment](https://gitea.phasefinal.com/vh/cicada) <sub>· brokkr-smithy-dev · 2026-09-20 14:11</sub>
+- [ShutterChute: denoise strength + EV lift on the 4 darkest Pancake Breakfast frames (1:1 crops)](http://10.100.10.50:8090/b/sc-denoise-ev/) <sub>· shutter-dev · 2026-09-20 15:12</sub>
+- [ShutterChute: DSC03888.ARW (ISO 12800, darkest frame) + current style — for authoring a working denoise in darktable](http://10.100.10.50:8090/b/sc-denoise-raw/) <sub>· shutter-dev · 2026-09-20 15:27</sub>
+- [Cutesy robot girl — 5 briefs x 2 seeds, 259-372 Hz, plus three robot textures (EVE / classic / WALL-E)](http://10.100.10.50:8090/b/robot-girl/) <sub>· tts-dev · 2026-09-20 15:53</sub>
+- [cicada-raw is LIVE — fastest voice on the fleet at 220.2 ms; reference + clones + the defect I retracted](http://10.100.10.50:8090/b/cicada-raw/) <sub>· tts-dev · 2026-09-20 16:06</sub>
+- [ShutterChute: denoise strength ladder on the REPAIRED split — 1:1 crops, 4 dark frames](http://10.100.10.50:8090/b/sc-denoise-strength/) <sub>· shutter-dev · 2026-09-20 16:24</sub>
+- [ShutterChute: four-way denoise comparison — no denoise / classical / SCUNet (automatable) / neural restore](http://10.100.10.50:8090/b/sc-denoise-fourway/) <sub>· shutter-dev · 2026-09-20 17:25</sub>
+- [ShutterChute: RawNIND UtNet2 pre-demosaic — 8.01 to 2.07 at 2.8s/frame, running outside darktable](http://10.100.10.50:8090/b/sc-rawdenoise/) <sub>· shutter-dev · 2026-09-20 18:47</sub>
+- [ShutterChute: frequency-selective detail recovery after raw AI denoise](http://10.100.10.50:8090/b/sc-detail-recovery/) <sub>· shutter-dev · 2026-09-20 18:54</sub>
+- [ShutterChute: raw AI denoise @70% across six frames, mean luminance 20 to 148](http://10.100.10.50:8090/b/sc-iso-spread/) <sub>· shutter-dev · 2026-09-20 18:58</sub>
+- [raw-denoise first real-model run: A raw vs B linear TIFF (black) vs C sRGB-encoded (tonality right, colour wrong)](http://10.100.10.50:8090/b/denoise-first-run/) <sub>· shutter-dev · 2026-09-21 06:45</sub>
+- [Pancake Breakfast low-light: raw vs denoised+2EV, full res + 1:1 crops; 3.3-3.5x noise reduction measured](http://10.100.10.50:8090/b/pancake-denoise/) <sub>· shutter-dev · 2026-09-21 07:02</sub>
+- [raw-denoise: TIFF handoff vs LinearRaw DNG handoff - the colour fix, before/after](http://10.100.10.50:8090/b/dng-handoff/) <sub>· shutter-dev · 2026-09-21 07:38</sub>
+- [EV ladder on a denoised Pancake frame: face luma vs frame median vs the 18% grey reference](http://10.100.10.50:8090/b/ev-ladder/) <sub>· shutter-dev · 2026-09-21 07:51</sub>
+- [golden-frame candidates for the one-and-done white balance: two lighting clusters, two each](http://10.100.10.50:8090/b/golden-candidates/) <sub>· shutter-dev · 2026-09-21 07:57</sub>
+- [Sindra @ 20 (v2, replaced) — 5 NSFW engines x 4 scenes x 2 seeds, 40 renders + 4 sheets + the age-lever diagnostic](http://10.100.10.50:8090/b/sindra20-engines/) <sub>· comfy-dev · 2026-09-21 07:57</sub>
+- [vibrance/saturation spike: 4 steps on a well-lit and a recovered frame; which colorbalancergb float is which, measured](http://10.100.10.50:8090/b/vibrance-spike/) <sub>· shutter-dev · 2026-09-21 08:29</sub>
+- [face metering measured on all 696 keepers: gate 20.7% -> 34.2%, 94 frames newly caught](http://10.100.10.50:8090/b/face-metering/) <sub>· shutter-dev · 2026-09-21 08:29</sub>
+- [darktable 5.6.1 on nh3-dev: the versions disagree, and the vibrance pick was made on 4.2.1](http://10.100.10.50:8090/b/dt56-recheck/) <sub>· shutter-dev · 2026-09-21 09:05</sub>
+- [Pancake Breakfast re-delivery: all 270 heroes, exposure + denoise + vibrance, SmugMug-ready](http://10.100.10.50:8090/b/pancake-v2-delivery/) <sub>· shutter-dev · 2026-09-21 09:48</sub>
+- [Draupnir — agent-directed parametric CAD for 3D printing; many harnesses propose, one gate decides](https://gitea.phasefinal.com/vh/draupnir) <sub>· brokkr-smithy-dev · 2026-09-21 10:47</sub>
+- [Pancake lift spike — Paula vs ours-zero-lift vs ours-metered, 8 frames](http://10.100.10.50:8090/b/pancake-lift-spike/) <sub>· shutter-dev · 2026-09-21 10:55</sub>
+- [Pancake dark band (face 17-42) — Paula vs ours at zero lift](http://10.100.10.50:8090/b/pancake-dark-band/) <sub>· shutter-dev · 2026-09-21 10:57</sub>
+- [Lift ladder — your 15 labelled frames at zero / +0.67 / +1.33 EV](http://10.100.10.50:8090/b/pancake-lift-ladder/) <sub>· shutter-dev · 2026-09-21 11:22</sub>
+- [Saturation+vibrance ladder — current / 75% / 50%, zero lift throughout](http://10.100.10.50:8090/b/pancake-saturation/) <sub>· shutter-dev · 2026-09-21 11:22</sub>
+- [Pancake v3 — the full 270 at cap 4/3, saturation 33%, gate/meter split](http://10.100.10.50:8090/b/pancake-v3-full/) <sub>· shutter-dev · 2026-09-21 13:29</sub>
+- [Pancake v3 — the 53 lifted frames vs Paula, worst blown first](http://10.100.10.50:8090/b/pancake-v3-lifted/) <sub>· shutter-dev · 2026-09-21 13:29</sub>
+- [Sigmoid colour test — Paula vs per-channel / RGB-ratio / smooth, 6 lifted + 2 unlifted controls](http://10.100.10.50:8090/b/pancake-sigmoid/) <sub>· shutter-dev · 2026-09-21 14:15</sub>
+- [Draupnir: 5 of 6 gate checks real — min-wall lands and the control pair finally separates (thin-wall FAILs at 1.0001mm vs 1.2mm floor); renders, STLs, calibration data](http://10.100.10.50:8090/b/draupnir-first-stl/) <sub>· draupnir · 2026-09-21 14:25</sub>
+- [Closed loop — 12 samples: Paula vs open-loop vs closed-loop, with EV and blown %](http://10.100.10.50:8090/b/pancake-closed-loop/) <sub>· shutter-dev · 2026-09-21 14:46</sub>
+- [Draupnir first commission: puck-light diffuser cap — 90.4mm shroud, 55.9% open, renders + STL/STEP (and the gate bug this part found)](http://10.100.10.50:8090/b/draupnir-puck-cap/) <sub>· draupnir · 2026-09-21 14:49</sub>
+- [Pancake v4 — the full 270 through the closed loop](http://10.100.10.50:8090/b/pancake-v4-full/) <sub>· shutter-dev · 2026-09-21 15:46</sub>
+- [Pancake v4 — the frames the loop changed, Paula / open / closed, worst blown first](http://10.100.10.50:8090/b/pancake-v4-changed/) <sub>· shutter-dev · 2026-09-21 15:46</sub>
+- [ShutterChute v0.9.14 on the mini — final triage over the 270 closed-loop deliveries](http://10.100.10.50:8477/?token=wtIRzaqmRQ3Qg2cjUwMZjd-OvNFB9UP3GXyGjDW2d-E&triage=/Users/paulahoang/Photos/PancakeBreakfast/deliver-shutterchute-260921) <sub>· shutter-dev · 2026-09-21 21:07</sub>
+- [ShutterChute v0.9.15 on the mini — triage, fit fixed](http://10.100.10.50:8477/?token=dG9y44XQmJfH7q8Wy_o2KKeIxGnUeP5zh8yADOoEYA4&triage=/Users/paulahoang/Photos/PancakeBreakfast/deliver-shutterchute-260921) <sub>· shutter-dev · 2026-09-21 21:50</sub>
+- [ShutterChute v0.9.16 — triage: centred delete tag, 1:1 pans](http://10.100.10.50:8477/?token=Nx9zEhTOXIfCrmpA72OEqEHGz7atxKQL8y5v3ecoW98&triage=/Users/paulahoang/Photos/PancakeBreakfast/deliver-shutterchute-260921) <sub>· shutter-dev · 2026-09-21 22:01</sub>
+- [cr123a-to-d-sleeve — renders, STL + STEP, gate WARN on the 0.8 mm shoulder](http://10.100.10.50:8090/b/cr123a-to-d-sleeve/) <sub>· draupnir · 2026-09-21 23:14</sub>
+- [Sindra @ 20 EVIDENCE BOARD — all 20 sheets + diagnostics, zero single frames (replaces the 122-image finalists board)](http://10.100.10.50:8090/b/sindra-evidence/) <sub>· comfy-dev · 2026-09-21 23:33</sub>
+- [infra-ops: five ERP run-7 decisions, open and unanswered since 2026-09-09](http://10.100.10.50:8090/b/run07-decisions/) <sub>· brokkr-smithy-dev · 2026-09-21 23:52</sub>
+- [Moody vs Realism BAKEOFF — 8 new scenes (4 SFW / 4 NSFW, no bedroom), 32 renders; verdict is a framing-dependent split](http://10.100.10.50:8090/b/sindra-bakeoff/) <sub>· comfy-dev · 2026-09-22 00:20</sub>
+- [krea2 LoRA portability test — RAW-trained LoRAs DO activate on distilled turbo checkpoints (3 seeds, null+negative+positive controls)](http://10.100.10.50:8090/b/krea2-lora-portability/) <sub>· comfy-dev · 2026-09-22 01:53</sub>
+- [The High Seat — SVOS board + Miranda (nh3-dev)](http://10.100.10.50:8770) <sub>· svos-dev · 2026-09-22 08:29</sub>
+- [Sindra training corpus pass 1 (54 frames) + the two validated fixes before the corrected re-render](http://10.100.10.50:8090/b/sindra-corpus-v1/) <sub>· comfy-dev · 2026-09-22 09:10</sub>
+- [Miranda re-minted Icelandic — 4 briefs x 2 seeds + Swedish/Norwegian discrimination controls + the incumbent](http://10.100.10.50:8090/b/miranda-is/) <sub>· tts-dev · 2026-09-22 10:41</sub>
+- [Sindra nude selection pool — 36 frames (10 rear), pick ~10 matching body shapes](http://10.100.10.50:8090/b/sindra-nude-pool/) <sub>· comfy-dev · 2026-09-22 11:08</sub>
+```

docs/contracts/u3_declared_embed_seam.contract.md

@@ -0,0 +1,471 @@
+---
+contract_version: "1.0"
+module: "booth.app (verbatim serving) + booth/static/embed.js"
+purpose: "A booth that ships its own index.html is the operator's most important surface -- his design reviews, his audition reports, his briefs -- and the Booth reaches into it with six regular expressions against arbitrary author HTML plus a placeholder DSL that substitutes rendered markup by pattern. Both work today and both are the single most fragile thing in the service. This unit replaces the whole class with a DECLARED SEAM: the page carries one line (`<script src=\"/_booth/embed.js\" defer></script>`), the Booth mounts its chrome through real DOM APIs, and a page that declares the line is served with ZERO Booth markup added to it. A page that does not declare it gets that one line appended at the end -- the only remaining mutation, and it needs no pattern matching at all. Operator ruling, 2026-09-21: the page declares itself, the Booth mounts into it."
+depends_on:
+  - "booth.marks (marks_for, open_marks, hold_read -- the pick records the payload renders. UNCHANGED by this unit: U3 changes how fragments REACH the page, never what a mark is. Read against booth/marks.py, not against the U2 contract's prose -- see the seam review.)"
+  - "booth/templates/_ask_inline.html (the `whole` / `question` / `submit` macros stay the ONE renderer of an ask fragment, called from the embed payload instead of from inject_asks. Its `styles()` macro is DELETED -- inject_asks was its only caller and the CSS moves into embed.js so the chrome is one asset. Macro signatures are otherwise untouched.)"
+  - "booth.asks.normalize_ask (TRANSITIVE, through `marks._hydrate`, and named because the payload shape depends on it: a MULTI ask normalizes to questions whose `key` matches `^[A-Za-z0-9][A-Za-z0-9._-]{0,60}$`, and a SINGLE-question ask normalizes to exactly one question whose `key` is `None`. Both facts are load-bearing -- the first makes splitting an anchor spec on the first colon unambiguous, the second is why `questions` is a list. Read against booth/asks.py:138-223.)"
+  - "booth.app.FAVICON_HREF (the data-URI icon, carried in the payload rather than copied into embed.js -- a third copy of that string is exactly the multiple-readers-of-one-truth shape the repo's ONE-RESOLVER rule (CLAUDE.md invariant 3 -- NOT this contract's INV-1, which is the untouched-page rule; a cold arm read the two as one label and was right to) exists to stop. base.html's literal copy predates this unit and is out of scope.)"
+language: "python + javascript"
+complexity: "medium"
+estimated_loc: 420
+confidence: 0.82
+used_by:
+  - "booth.app.booth_view (the verbatim branch: one read, one substring check, one conditional append -- replacing inject_asks + wrap_verbatim_html entirely)"
+  - "the operator's verbatim reports (4 of 21 live booths ship their own index.html; 2 of those 4 use the placement DSL, so the migration is not hypothetical)"
+  - "report authors (the declared line is the new public API for a booth that wants Booth chrome where it chooses)"
+touches:
+  - "booth/static/embed.js (NEW -- the mount script and the chrome CSS, one asset. Read ONCE at app startup, never per request; see INV-5.)"
+  - "booth/app.py (DELETE wrap_verbatim_html, _insert_before, _insert_after, _ICON_RE, _HEAD_CLOSE_RE, _HTML_OPEN_RE, _DOCTYPE_RE, _BODY_CLOSE_RE, _HTML_CLOSE_RE, _BACK_CHIP, asks_chip, inject_asks, FAVICON_LINK and the `from booth.inline import` block. ADD EMBED_SRC/EMBED_SCRIPT_TAG, the startup read of embed.js, GET /_booth/embed.js, GET /b/{name}/embed.json, and the rewritten verbatim branch of booth_view.)"
+  - "booth/inline.py (DELETED ENTIRELY -- 119 lines. Nothing else imports it; `scripts/booth` never did, so the stdlib-only CLI surface is untouched. ONE line survives the module: `form_id`, which builds the shared form element id the fragments bind to, moves into booth/app.py beside the route that renders them. It is not placement machinery and dying with the placement engine would take the fragments with it. Seam review, SR-3.)"
+  - "booth/templates/_ask_inline.html (DELETE the `styles()` macro and rewrite the header comment: the fragments are now mounted by embed.js, not substituted by regex, and `No JavaScript` stops being true.)"
+  - "tests/test_booth.py (DELETE the five test_wrap_* tests and test_verbatim_booth_wrapped_with_back_chip -- they test a mechanism this unit removes; the FAVICON_LINK import goes with them)"
+  - "tests/test_asks.py (the inline-placement block, ~L480-590: assertions that server-rendered fragments appear in the page body become assertions about the embed payload. The BEHAVIOUR they encode -- every question reachable, a scattered form still submittable, a typo'd id left alone -- is preserved and re-asserted, half in Python and half in the browser.)"
+  - "tests/test_embed.py (NEW -- the payload, the injection rule, the no-hot-reload guard)"
+  - "tests/test_embed_browser.py (NEW -- Playwright against a real Chromium: the placement algorithm and the form association, neither of which the Python suite can see. SKIPS, never fails, when playwright or the shared browser is unavailable.)"
+  - "pyproject.toml (test extra gains `playwright>=1.60,<1.63` -- the range is the set of releases whose pinned Chromium revision is present in the box-wide /opt/ms-playwright store. Stated explicitly because it is invisible otherwise: 1.63 wants chromium-1243, which is NOT there, and the failure is an opaque `Executable doesn't exist`.)"
+  - "docs/design/information-architecture.md (the `What this deletes` list becomes what this DID delete; the standalone /asks bullet is corrected -- U2 already reduced it to a 308)"
+  - "ROADMAP.md (U3 row struck through; the ordering table gains the two orders this unit states)"
+assumptions:
+  - "THE OPERATOR ALREADY RULED ON THE SEAM (2026-09-21, recorded in the IA doc): the page declares itself and the Booth mounts into it, via `<script src=\"/_booth/embed.js\" defer></script>`. That ruling ACCEPTS a JavaScript dependency on the verbatim path, which today has none. This contract does not re-open it. What the contract DOES do is state the consequence plainly so it is not discovered later -- see the degradation assumption below."
+  - "THE ONLY REMAINING SERVER-SIDE MUTATION IS A CONDITIONAL APPEND, AND IT NEEDS NO PATTERN AT ALL. Two substring tests (`src=\"/_booth/embed.js\"` and its single-quoted twin), then a concatenation. ⚠ THE BARE PATH WAS THE FIRST DRAFT AND IT FAILED IN THE DANGEROUS DIRECTION: a report that merely MENTIONS the path -- a code sample, a comment, a sentence about this feature, which the Booth's own design reports are the likeliest pages to contain -- would have counted as declaring it, been served untouched, and shown no chrome at all, silently. Requiring `src=` immediately before the path flips the failure direction: an unusual spelling (`src = \"...\"`, an unquoted attribute, a `?v=2` suffix) reads as NOT declared, so a second tag is appended and embed.js mounts once anyway on its `window.__boothEmbed` guard. A missed declaration costs a duplicate tag; a false one costs the operator his chrome. Three of four cold-panel arms found this independently. This is why all six regexes die rather than collapsing to one: content appended AFTER `</html>` is parsed into the body by every browser, so there is nothing to find. Nothing is ever PREPENDED, which is what retires both of wrap_verbatim_html's hard constraints in one stroke -- no doctype can be displaced into quirks mode and no charset meta can be pushed out of the first 1024 bytes, because nothing moves."
+  - "THE FAVICON MOVES FROM A REGEX TO A DOM QUERY. `_ICON_RE` existed to answer `does this page already declare an icon`, against raw text, and three more regexes existed to find a head-ish seam to put one in. embed.js asks `document.querySelector('link[rel~=\"icon\"]')` and appends to `document.head`. That is the same question and the same action, asked of a parsed document instead of a string -- and it is four of the six regexes."
+  - "FRAGMENTS ARE STILL RENDERED BY JINJA, ONLY PLACED BY JAVASCRIPT. The payload carries server-rendered HTML from the EXISTING `_ask_inline.html` macros. Re-implementing the ask form in JavaScript would make two renderers of one truth, which is precisely the shape the repo's ONE-RESOLVER rule (CLAUDE.md invariant 3) was written to stop after the zoom view lost its captions. embed.js does DOM placement and nothing else: it never decides what a mark says, whether it is open, or what order marks come in."
+  - "PLACEMENT IS AN ANCHOR-FILL, NOT A REPLACEMENT, AND THAT IS A DELIBERATE CHANGE FROM TODAY. `_EL_RE` matches an author's opening tag and SUBSTITUTES it, so `<div class=\"ask\" data-booth-ask=\"dfa:logo\"><h3>The one asset that must survive</h3>` loses both the wrapper's class and -- visually -- its framing, leaving the author's heading orphaned and the closing `</div>` stray. That is live today on `dfa-concepts`. embed.js uses `el.insertAdjacentHTML('beforeend', frag)`: the author's element and its contents survive and the fragment lands inside, under the heading. Strictly closer to what the markup says, and it is the behaviour a DOM API gives for free."
+  - "`data-booth-mark` IS CANONICAL; `data-booth-ask` IS A KEPT ALIAS. U2 made an ask one shape of mark and the IA doc names the anchor `data-booth-mark`. But 2 of the 4 live verbatim booths use the `data-booth-ask` spelling, in the operator's own reports, so the selector accepts both -- one extra clause in one selector string. Same for `data-booth-mark-submit` / `data-booth-ask-submit`. Renaming without the alias would break a live report to save nothing."
+  - "THE HTML-COMMENT PLACEHOLDERS ARE DROPPED, NOT PORTED. `<!-- booth:ask stem -->` and `<!-- booth:ask-submit stem -->` have ZERO users across all 21 live booths. Walking comment nodes to keep them would be real complexity bought for nobody, in the unit whose entire point is deletion. A page that used one degrades to the append path -- the ask still renders, at the end -- so the never-invisible guarantee holds even for a caller we do not know about."
+  - "DEGRADATION WITH JAVASCRIPT OFF IS A REAL LOSS AND IT IS NAMED HERE. Today the verbatim path is zero-JS: an ask renders server-side and submits through a plain form. After this unit, no JS means no chrome on the report -- no ask, no way home, no icon. The guarantee that an ask is NEVER INVISIBLE survives in a weaker and still-true form, through surfaces that need no script: the index card carries the open-mark badge, and `/b/<name>/marks` renders every mark server-side. This is the cost of the operator's ruling, stated once so nobody meets it as a surprise."
+  - "EMBED.JS IS READ ONCE AT STARTUP, FOR THE REASON TEMPLATES ARE. Serving it from disk per request would give the service a third staleness rule, and a live asset editable under a running process is exactly what put 19 of 25 booths at 500 on 2026-09-21. One rule in this repo: nothing takes effect until you restart. INV-5 holds the line the same way `test_templates_do_not_hot_reload_from_disk` does."
+  - "THE PAYLOAD ENDPOINT DOES NOT RECORD A VIEW. `booth_view` already calls `record_view` above both early returns (U4), and `.viewed` is a deliberate look. A fetch issued by a script on a page that has ALREADY been recorded would double-count activity and reset the TTL on machinery rather than on the operator -- the same distinction the `.lock` exemption draws in `_newest_mtime`."
+  - "THE READ IS LENIENT AND THE STATUS STAYS 200, copied deliberately from `/b/{name}/marks.json`. A damaged `.marks.json` must not 500 the operator's report; it returns an `error` in the body and embed.js mounts the nav anyway. This is the v0.2.2 lesson and the posture every read path in this service already takes."
+  - "WRAP_MAX_BYTES SURVIVES UNCHANGED, at 8 MiB, with the same raw-FileResponse fallback. The work behind it is now trivial, but the READ is not: the largest live verbatim booth is 280 KB and a pathological one still should not be pulled into memory. A booth over the cap loses its chrome exactly as it does today -- no regression, and the constant keeps its existing test."
+open_questions:
+  - "Whether `/_booth/embed.js` should eventually carry the gallery page's chrome too, making one embed for both surfaces. Out of scope: the gallery page is server-rendered end to end and has no seam problem to solve."
+  - "Whether a booth should be able to suppress injection entirely (a `.no-embed` dotfile) for a report that wants to be served truly untouched. No live booth wants it; declaring the line and then not using it is already most of the way there. Parked rather than designed."
+---
+
+# U3 — the declared embed seam
+
+## The defect, stated precisely
+
+A booth that ships its own `index.html` is served verbatim. That is the whole
+promise of the verbatim path, and the Booth breaks it twice on the way out:
+
+1. **`wrap_verbatim_html`** searches arbitrary author HTML with six regular
+   expressions — `_ICON_RE`, `_HEAD_CLOSE_RE`, `_HTML_OPEN_RE`, `_DOCTYPE_RE`,
+   `_BODY_CLOSE_RE`, `_HTML_CLOSE_RE` — to find somewhere to put a favicon and
+   somewhere to put a floating chip, while threading two constraints it cannot
+   verify: never put anything ahead of a leading doctype, and keep the charset
+   meta inside the first 1024 bytes.
+2. **`booth/inline.py`** matches a placeholder DSL with four more patterns and
+   substitutes rendered HTML into the author's markup by string replacement.
+
+Ten patterns, applied to documents the Booth did not write, does not parse, and
+cannot validate. It works. It is also the single most fragile thing in the
+service, and it is load-bearing for the operator's most important workflow.
+
+The failure this invites is not a crash. It is a report that renders *slightly*
+wrong — and there is a live specimen already. `dfa-concepts/index.html` writes:
+
+```html
+<div class="ask" data-booth-ask="dfa:logo"><h3>The one asset that must survive</h3>
+```
+
+`_EL_RE` matches the opening `<div …>` and replaces it. The author's `.ask`
+wrapper class is gone, the `<h3>` is orphaned, and the `</div>` further down is
+stray. Nobody filed a bug, because a page that is 95% right does not look broken.
+
+## The seam
+
+Operator ruling, 2026-09-21. A report carries one line:
+
+```html
+<script src="/_booth/embed.js" defer></script>
+```
+
+and the Booth mounts its chrome through real DOM APIs. Three consequences, and
+the third is the one worth stating out loud:
+
+- **A page that declares the line is served with nothing added to it.** Not
+  "one small injection" — nothing. The body is what the author wrote.
+- **A page that does not declare it gets that one line appended at the end.**
+  A substring test and a concatenation; no pattern, nothing prepended, no
+  constraint to thread.
+- **Both of `wrap_verbatim_html`'s hard constraints stop existing** rather than
+  being satisfied more carefully. You cannot displace a doctype you never move,
+  and you cannot push a charset meta out of the detection window by appending.
+
+## What crosses the seam
+
+`GET /b/{name}/embed.json` — server-rendered fragments, and nothing embed.js has
+to decide for itself:
+
+```json
+{
+  "booth": "dfa-concepts",
+  "home": "/",
+  "favicon": "data:image/svg+xml,…",
+  "open": ["dfa"],
+  "marks": [
+    {
+      "id": "dfa",
+      "error": null,
+      "whole":  "<div class=\"bk-ask\" …>",
+      "submit": "<div class=\"bk-ask\" …>",
+      "questions": [
+        {"key": "logo",    "html": "<div …>"},
+        {"key": "display", "html": "<div …>"}
+      ]
+    }
+  ]
+}
+```
+
+Every HTML string comes from the `_ask_inline.html` macros that render the same
+fragments today. `open` is `open_marks(picks)` — computed once, server-side, and
+never re-derived in JavaScript.
+
+**A payload whose `.marks.json` could not be read has a stated shape**, because
+an arm asked and the first draft did not say: `marks` is `[]`, `open` is `[]`,
+`booth` / `home` / `favicon` are present as normal, and top-level `error` and
+`detail` carry the verdict. Status stays 200, copied from `/marks.json` — a
+pinned status code is a promise to remote clients, and the information goes in
+the body instead. The nav mounts; nothing else does. A per-mark `error` is a
+different thing: that is ONE unreadable pick inside a file that read fine.
+
+**`questions` is a LIST, and `key` may be `null`.** This is not a style choice.
+A single-question pick normalizes to `questions: [{"key": None, …}]`
+(`asks.normalize_ask`), so a JSON object keyed by question key would serialize
+that key as the string `"null"` — inventing a name that collides with a real key
+and that JavaScript would have to translate back. A list also carries declaration
+order in the format itself rather than leaning on object-key insertion order.
+A `null` key matches no anchor spec, which is correct and is exactly what
+`place` does today: a single-question pick is addressed as a whole or not at all.
+Found by the seam review; see SR-2.
+
+## How the script learns which booth it is on
+
+**The find of the contract-review round, three arms independently**, and the one
+gap that made the rest unimplementable as first written: the declared line is
+byte-identical for every booth, the payload endpoint needs `{name}`, and the
+name arrives *inside* the response the fetch needs the name to make.
+
+The rule, stated once:
+
+> **The booth name is the second path segment of the page's own address.** A
+> verbatim report is served at `/b/<name>/`, so `embed.js` reads
+> `location.pathname`, takes segment 2, and `decodeURIComponent`s it. A page
+> whose address is not `/b/<name>/...` mounts nothing and returns quietly.
+>
+> **Override:** a `<script data-booth="...">` attribute wins if present. The
+> Booth never writes one — the appended tag is exactly the documented line — but
+> an author embedding a report elsewhere needs a way to say so, and one optional
+> attribute is cheaper than a second endpoint.
+
+This makes the URL grammar part of the public seam, which is the honest reading:
+an author who writes the line is relying on where the Booth serves them, and
+that should be written down rather than inferred.
+
+## The placement algorithm
+
+The same algorithm `inject_asks` runs today, expressed against a DOM instead of
+a string. It is written out here because it is the part that moves languages,
+and a reviewer has to be able to check the two against each other.
+
+```
+placed    : Map<markId, Set<key | WHOLE>>
+submitted : Set<markId>
+
+1. every element matching
+      [data-booth-mark], [data-booth-ask]          -- in document order
+   spec -> (id, key?)  by splitting on the first ":"
+   mark unknown            -> leave the element ALONE (a typo stays visible)
+   key absent              -> mount whole;      placed[id] += WHOLE; submitted += id
+   key names no question   -> leave the element ALONE
+   key present             -> mount question;   placed[id] += key
+
+2. every element matching
+      [data-booth-mark-submit], [data-booth-ask-submit]
+   mark unknown            -> leave alone
+   otherwise               -> mount submit;     placed[id] ||= {}; submitted += id
+
+3. tail, appended to <body> in payload order. THE ARROWS ARE EXCLUSIVE, NOT
+   CUMULATIVE -- first match wins and the mark is done. An arm read them as
+   four independent tests, under which one unplaced mark would mount its whole
+   form AND every question AND a submit block; the notation allowed it and the
+   prose did not forbid it:
+
+   if id not in placed:            append whole; NEXT MARK
+   elif mark.error:                append nothing; NEXT MARK
+   else:
+       if WHOLE not in placed[id]: append every question not in placed[id]
+       if id not in submitted:     append submit   (scattered, still submittable)
+
+4. re-associate: for every control carrying form="…", remove and re-set the
+   attribute, so its form owner is resolved after all fragments are in place.
+
+5. chip: if `open` is non-empty, link it to the FIRST element in document order
+   whose id is EXACTLY `bk-ask-<open[0]>` or begins `bk-ask-<open[0]>-`.
+   A bare prefix match would send the chip to `bk-ask-batch2-r1` for the mark
+   `batch`, or to an author's own element -- flagged by a cold arm, and the
+   trailing hyphen is what rules it out.
+
+Two more rules the first draft left to the selector rather than stating:
+
+- **An element carrying BOTH `data-booth-mark` and `data-booth-ask` uses the
+  canonical one.** The alias exists for reports written before the rename, not
+  to double a mount.
+- **A submit anchor's spec is its stem; any `:key` on it is IGNORED.** There is
+  no per-question submit block — one pick has one `<form>`, which is the whole
+  reason the `form=` binding exists.
+```
+
+**`mount` is `el.insertAdjacentHTML('beforeend', frag)`** — the anchor element
+and its existing contents survive; the fragment lands inside. See the assumption
+on anchor-fill for why this is a deliberate change and not an accident.
+
+**Step 4 is measured, not assumed.** Chromium 151 resolves a control's form owner
+correctly even when the control is inserted before its `<form>`: a probe run
+2026-09-22 (N=3 per condition, with a form-first positive control and a
+points-at-nothing negative control) returned `F, F, F` for control-first and
+`null, null, null` for the negative. So the pass is *not* needed in Chromium.
+It is three lines, it costs nothing, and the sensitivity floor of that probe is
+**one engine** — the operator's own browser was not measured. The failure it
+guards against is a form the operator fills in whose controls reach no form,
+so the button does nothing.
+
+**Step 5 deletes an element.** Today `inject_asks` injects `<a id="bk-ask-<id>-top">`
+before the first fragment of each pick so the chip has somewhere to jump. The
+fragments already carry ids; document order in a live DOM is directly queryable;
+the extra anchor is not needed.
+
+## Invariants
+
+Each is falsifiable by a change that a test must catch going red. The
+*Falsifiable:* line names that change — not a test that merely mentions the
+invariant. (Five of seven U4 falsifiers were vacuous; see
+`persistent-memory.d/2026-09-22-vacuous-falsifiers.md`.)
+
+**INV-1 — A page that declares the seam is served BYTE FOR BYTE.**
+The response body for a verbatim booth whose `index.html` contains
+`src="/_booth/embed.js"` (either quote style) is exactly the bytes on disk.
+⚠ **Bytes, not text, and that is a correction.** The first implementation read
+with `read_text()`, which opens in universal-newline mode: a CRLF report came
+back LF, and `errors="replace"` turned any non-UTF-8 byte into U+FFFD. A
+declaring page was NOT served as its author wrote it — the headline promise —
+and the test could not see it, because its fixture was LF-only ASCII. The file
+is decoded only to ask whether it declares the seam; what goes on the wire is
+the original bytes. A page that only mentions the path is NOT declaring it — see the
+conditional-append assumption for which way that has to fail.
+*Falsifiable:* append anything — a chip, a comment, a newline — to the declaring
+branch's response and `test_declaring_page_is_served_untouched` fails on a
+whole-body equality, not on a substring absence.
+
+**INV-2 — A page that does not declare the seam, AND IS UNDER `WRAP_MAX_BYTES`,
+is mutated exactly once, at the end.** The response is the source BYTES plus
+`EMBED_SCRIPT_TAG`'s bytes and nothing else, with the source a byte-exact
+prefix of it.
+⚠ **The size cap is an explicit exception, not an oversight** — two cold arms
+read the invariant's universal wording against the raw-`FileResponse`
+assumption and found them prescribing different responses for the same page. An
+over-cap page is mutated ZERO times and loses its chrome, exactly as it did
+before this unit.
+*Falsifiable:* insert the tag before `</head>` instead of appending, or add the
+favicon link back, and `test_undeclared_page_gains_only_the_tag` fails the
+prefix assertion. The exception has its own test,
+`test_an_oversize_verbatim_page_is_served_raw`, which fails if the append starts
+firing above the cap.
+
+**INV-3 — No regular expression is applied to author HTML.**
+The verbatim branch of `booth_view` performs two `in` tests and one `+`.
+⚠ **The first draft of this falsifier was VACUOUS and three arms caught it.**
+It name-matched the six deleted patterns, so reintroducing the same regex under
+a new name — `_TAIL_RE`, applied in the verbatim branch — left the test green,
+on this contract's central promise. Worse, this repo's own vacuity pass missed
+it, because the mutation it tried was the named one: **a vacuity pass is only as
+good as the mutation it picks, and picking the one the contract names is how it
+agrees with itself.**
+*Falsifiable:* `test_no_regex_touches_author_html` walks the AST of
+`booth/app.py` and asserts the module performs **exactly one** regex operation
+— `ask_form_id`'s `re.sub` over a mark id, which is not a page — plus that
+`booth/inline.py` does not exist. Any regex anywhere in the module, under any
+name, fails it. Verified by mutation: a renamed `_TAIL_RE.sub` in
+`embed_verbatim` goes red, and the unmutated control stays green.
+
+**INV-4 — The payload is the only source of what a mark says.**
+embed.js never decides openness, order, or content. `open` comes from
+`open_marks`; `marks` order is `marks_for` order; `questions` order is
+declaration order.
+*Falsifiable:* the claim ranges over three things and so does the check.
+**Openness:** have embed.js derive open marks from a `bk-done` class and
+`test_the_chip_count_comes_from_the_server` fails on a half-answered pick, which
+`open_marks` calls open and the rendered state does not. **Order:** reverse the
+tail iteration and `test_the_tail_follows_payload_order` fails. **Content:** the
+fragments are strings the page never authors, which
+`test_every_piece_the_author_can_place_is_offered` pins on the server side.
+
+**INV-5 — `/_booth/embed.js` is read once at startup.**
+*Falsifiable:* change the route to `read_text()` per request and
+`test_embed_js_does_not_hot_reload_from_disk` fails — it mutates the file on
+disk after the app is built and asserts the served body is unchanged.
+
+**INV-6 — Every ordered collection this unit renders has a stated rule.**
+Anchors are visited in **document order** (`querySelectorAll`). The tail is
+appended in **payload order**, which is `(created, id)` — the rule `marks_for`
+and `hold_read` both sort by, stated here as the rule rather than as one
+function's name. Questions
+within a mark are in **declaration order**. The chip targets the **first element
+in document order** whose id starts with the open mark's prefix.
+*Falsifiable:* sort the tail by anything else — id, key, insertion — and
+`test_tail_order_is_payload_order` fails against a fixture whose creation order
+and id order disagree.
+
+**INV-7 — Every question of every READABLE pick reaches the document, on a
+page that runs the script.** Either placed at an anchor or appended, and every
+pick with a placed question has a submit block.
+⚠ **Two qualifiers, both added because arms read the first wording literally and
+were right.** *Readable*: a pick carrying `error` has no questions to place —
+`marks._hydrate` gives it an empty list — so the tail mounts its broken-ask box
+and stops, and an unqualified "every pick" would have demanded placement the
+algorithm forbids in exactly the damaged-data case the leniency posture exists
+for. *Reaches the document*, not "is visible": the Booth cannot police an author
+who hides their own anchor, and a guarantee that claimed to would be unenforceable
+rather than strict.
+*Falsifiable:* drop the "append the questions the author did not place" branch
+and `test_partially_marked_page_still_shows_every_question` fails in the browser
+with 2 of 4 radio groups present.
+
+## Out of scope (deferred or never)
+
+Named so a reviewer does not read them as drift.
+
+- **The gallery page's chrome.** Only a booth's own `index.html` is served
+  verbatim; every other surface is server-rendered end to end and has no seam
+  problem. `/_booth/embed.js` is not loaded there and is not meant to be.
+- **Re-rendering an ask in JavaScript.** The payload carries server-rendered
+  HTML and embed.js places it. A JS renderer would be a second renderer of one
+  truth — the bug the repo's one-resolver rule exists to stop.
+- **A no-JavaScript fallback on the verbatim path.** The operator's 2026-09-21
+  ruling accepts the script dependency. The never-invisible guarantee degrades
+  to surfaces that need no script (the index card's badge, `/b/<name>/marks`),
+  and that is the stated cost, not an oversight to be fixed here.
+- **The HTML-comment placeholders** `<!-- booth:ask … -->`. Zero users across
+  all 21 live booths; dropped rather than ported. A page that used one falls
+  back to the append path, so its ask still renders.
+- **`_ask_inline.html`'s dead `standalone=False` macro parameter.** No caller
+  has passed `True` since U2 turned the standalone asks page into a 308.
+  Deleting it is tidy-up and changes a macro signature for no behavioural gain.
+- **`base.html`'s literal duplicate of the favicon data URI.** It predates this
+  unit. The payload reads `FAVICON_HREF`, so this unit adds no third copy; it
+  does not remove the second.
+- **`WRAP_MAX_BYTES` and its raw-serve fallback.** Unchanged at 8 MiB. A booth
+  over the cap loses its chrome exactly as it did before — no regression, and
+  the constant keeps its existing test.
+- **`GET /b/<name>/asks`.** Already a 308 into `/marks` since U2. Left alone:
+  the URL is in the operator's history and in landed reports.
+- **Pushing, and the version bump tier.** Minor needs the operator's approval.
+
+## Slices
+
+| # | slice | red→green on |
+|---|---|---|
+| 1 | `GET /b/{name}/embed.json` — payload shape, order, leniency, no view recorded | payload tests; existing 410 stay green |
+| 2 | `GET /_booth/embed.js` — served from a startup read, ETag, no hot reload | INV-5 |
+| 3 | the verbatim branch rewritten; `inject_asks` and `wrap_verbatim_html` deleted | INV-1, INV-2, INV-3 |
+| 4 | `booth/static/embed.js` — nav, favicon, styles, no marks yet | browser: chip present, icon set, declaring page untouched |
+| 5 | placement: anchors, tail, submit, re-association | browser: INV-4, INV-6, INV-7; the live `dfa-concepts` and `sindra-voice-1` shapes as fixtures |
+| 6 | delete `inline.py`; retire the six tests that test the deleted mechanism; docs | suite green, IA doc and ROADMAP updated |
+
+## Seam review
+
+The sibling-aware pass, run in-session against the real module surfaces rather
+than against the sibling contracts' prose. `/heid-contract-review` is
+artifact-only by design and structurally cannot see `booth/marks.py`, so this is
+the only gate that can check what the contract borrows from it.
+
+| # | finding | disposition |
+|---|---|---|
+| **SR-1** | The order invariant named `marks_for`'s ordering. The route actually reads through `hold_read` — one read answering both "what is here" and "can it be read", per the TOCTOU lesson — and only falls back to `marks_for` on the error path. Both sort `(created, id)`, so the contract was not wrong, but it named a function where it meant a rule. | **Amended.** INV-6 states the rule. The route's reader is named in the payload section. |
+| **SR-2** | **The payload shape was wrong.** `questions` as a JSON object keyed by question key breaks on a single-question pick, whose only question has `key: None` (`asks.normalize_ask`, the `multi: False` branch) — `json.dumps` writes that key as the string `"null"`. Every one-question ask in the fleet hits it, including the live `sindra-voice-1`. | **Scope fix.** `questions` is a list of `{key, html}`; `key` is nullable; declaration order is carried by the format. `booth.asks.normalize_ask` added to `depends_on`. |
+| **SR-3** | `inline.form_id` was inside the module the contract deletes entirely, but it is not placement machinery — it builds the shared `<form>` id the question fragments bind to with `form=`. Deleting the module as written would delete the fragments' ability to submit. | **Scope miss.** `form_id` moves to `booth/app.py`; `touches` says so. |
+| **SR-4** | A FLAG mark's id is literally `flag:<target>` (`marks.flag_id`) — it contains the separator the anchor spec splits on. It never reaches the payload only because the payload filters `shape == "pick"`, and pick ids are `valid_stem`-checked (no colon). | **No change, stated.** The filter is load-bearing, not incidental; a later widening of the payload to all shapes would break the split rule silently. |
+| **SR-5** | `_ask_inline.html`'s `question(a, q, form_id, name_url, standalone=False)` has had no caller passing `standalone=True` since the standalone asks page became a 308 in U2. Dead parameter on a macro this unit edits. | **Out of scope, noted.** Deleting it is tidy-up, not this unit's work, and it changes a macro signature for no behavioural gain. |
+
+## Contract review — the cold panel
+
+`/heid-contract-review`, four arms, dispatched `01M351WKV666D681SSRNY7D7X6`.
+Triaged per the cross-frontier discipline: adopted on merits, not on authority.
+
+| # | finding | arms | disposition |
+|---|---|---|---|
+| **CR-1** | **The seam never tells `embed.js` which booth it is on.** The declared line is byte-identical for every booth, the payload endpoint needs `{name}`, and the name arrives inside the response the fetch needs it to make. Every other section depends on this unstated hop. | 3 of 4, independently | **Genuine add, and the round's headline.** The code already derived it from `location.pathname`; the CONTRACT did not say so, which makes a "public API" whose discovery mechanism is unspecified not fully one. New section: *How the script learns which booth it is on*. No code change. |
+| **CR-2** | **INV-3's falsifier was vacuous** — it name-matched the six deleted patterns, so a renamed regex applied to the page body kept it green, on this contract's central promise. | 3 of 4 | **Genuine add, and a CODE-side fix.** The test now asserts `booth/app.py` performs exactly one regex operation anywhere in the module. Verified by mutation in both directions. The lesson is sharper than the fix: **this repo's own vacuity pass missed it because it tried the mutation the contract named** — a pass that picks the named mutation agrees with itself. |
+| **CR-3** | **Declaration by bare substring fails in the dangerous direction.** A report that merely mentions `/_booth/embed.js` — a code sample, a comment — counted as declaring it and was served with no chrome at all, silently. | 3 of 4 | **Genuine add, CODE-side.** Detection now requires `src="…"` (either quote style), which fails toward a harmless duplicate tag instead. New test covers prose, comment and `?v=2` spellings. |
+| **CR-4** | **INV-2 and the size cap prescribe different responses** for an over-cap non-declaring page, and neither the invariant's wording nor a named falsifier carved the exception. | 2 of 4 | **Genuine add.** INV-2 now states the cap as an explicit exception and names the test that holds it. Code and test were already right. |
+| **CR-5** | **The tail's four arrows read as independent tests**, under which one unplaced mark mounts its whole form AND every question AND a submit block. | 1 | **Genuine add.** The notation allowed it and the prose did not forbid it. The block is now explicit if/elif/else. Code was already exclusive. |
+| **CR-6** | **INV-7 quantified over picks the algorithm filters** (errored picks) and over "visible", which placement cannot guarantee. | 2 of 4 | **Genuine add, wording.** INV-7 is now scoped to READABLE picks and claims *reaches the document*, not *is visible*. |
+| **CR-7** | The chip's prefix rule can select `bk-ask-batch2-r1` for mark `batch`, or an author's own element. | 1 | **Sharpening.** The code always matched exactly-or-hyphen; the contract said "starts with". Wording fixed, and `test_the_chip_does_not_jump_to_a_mark_that_merely_shares_a_prefix` now holds it. |
+| **CR-8** | Precedence undefined when one element carries both attribute spellings; submit-anchor key handling unstated. | 1 | **Sharpening.** Both stated; `test_the_canonical_attribute_wins_when_both_are_present` added. |
+| **CR-9** | The damaged-`.marks.json` payload shape was never stated — per-mark `error` was the only error shown. | 1 | **Genuine add, wording.** Stated in *What crosses the seam*. Test already existed. |
+| **CR-10** | "INV-1" names two different obligations — this contract's untouched-page rule, and the repo's one-resolver rule the assumptions cite. | 1 | **Genuine add, wording.** The assumptions now name CLAUDE.md invariant 3 explicitly. A real collision: the local falsifier goes red on an added newline and stays green if embed.js becomes a second renderer. |
+| **CR-11** | INV-4's falsifier covered openness while the invariant claimed openness, order AND content. | 1 | **Sharpening.** The falsifier now names a test per clause. |
+| **CR-12** | `html.questions` keyed by question name vs the top-level `questions` list — which is authoritative? And INV-4 naming `marks_for`'s order while INV-6 fixed `(created, id)`. | 2 | **Settled before the reply landed.** The in-session seam review collapsed both (SR-1, SR-2) while the panel was in flight. Independent convergence on the same two spots — worth recording, not re-fixing. |
+
+**One arm's finding not adopted**, and the reason: that a question mounted into
+an author-hidden anchor is still invisible. True, and out of reach — the Booth
+cannot police an author hiding their own markup. Answered by narrowing INV-7's
+claim rather than by chasing actual visibility (CR-6).
+
+**Methodology note the panel raised on its own**, relayed by heid: 5 of 8 arms
+across two unrelated callers the same evening independently proposed promoting
+the end-to-end seam-walk from a conditional deliverable to a mandatory one.
+CR-1 is a direct product of that exercise. Recorded here as evidence; the skill
+change is the operator's call, not this repo's.
+
+## Bug hunt — the cold panel
+
+`/heid-bug-hunt`, four arms, artifact-only over the merge-base diff, dispatched
+`01M352TPCSN52G6NGJ07T5WSGY`. ⚠ **The snapshot predates the contract-review
+fixes**, so two of its findings were already closed when the reply landed; the
+arms flagged the staleness themselves.
+
+| # | finding | arms | disposition |
+|---|---|---|---|
+| **BH-1** | **A declaring page was NOT served as written.** `read_text()` opens in universal-newline mode, so a CRLF report came back LF, and `errors="replace"` replaced any non-UTF-8 byte. The headline promise, broken by the read itself — and invisible to a test whose fixture is LF-only ASCII. | 1 | **Genuine add, and the best finding of the round.** The verbatim branch reads and serves BYTES; the decoded copy answers only "does it declare?". INV-1 and INV-2 now state the byte-level promise, with a CRLF-plus-invalid-byte fixture. |
+| **BH-2** | **A submit anchor inside the author's own `<form>` loses ours** — the HTML parser drops a nested form outright. Every control's `form=` then points at nothing, and the code recorded the pick as submitted so the tail added no fallback. The operator fills it in and the button does nothing. | 1 | **Genuine add.** A submit anchor counts as submitted only if the form actually survived (`hasForm`); otherwise the tail supplies one at body level, where no form encloses it. |
+| **BH-3** | **A broken pick's diagnostic never rendered from a submit-only anchor.** An errored pick's `submit` is empty; mounting that and marking it placed made the tail skip it, so the "broken ask" box vanished from the one surface built to show it. | 3 of 4 | **Genuine add.** A submit anchor for an errored pick is left alone, exactly as an anchor naming no mark is, and the tail mounts the diagnostic. |
+| **BH-4** | **An author's own element can hijack the chip.** `<section id="bk-ask-winner-background">` satisfies any id-prefix rule — the hyphen boundary from CR-7 included. | 4 of 4 | **Genuine add, and it supersedes CR-7's fix.** The chip now searches only the elements THIS SCRIPT MOUNTED, which is the identity the deleted `bk-ask-<id>-top` anchor used to guarantee, and takes the earliest of those by `compareDocumentPosition`. |
+| **BH-5** | **No error boundary around fragment rendering.** A `.marks.json` that is well-formed JSON with a wrong-shaped `answer` hydrates with no error and then raises in the macro. | 1, `needs-repro` | **Genuine add — reproduced before building for it.** `_safe_fragments` returns a per-mark error record, the same leniency `_hydrate_safe` applies one layer down. ⚠ **The gallery and marks pages still 500 on it, and that is PRE-EXISTING** — measured at `42ea67f`. Out of scope here and recorded rather than quietly widened: `persistent-memory.d/2026-09-22-a-wrong-shaped-answer-500s-the-gallery.md`. **CLOSED 2026-09-22**, after U6, at the hydration boundary rather than by a third copy of this guard — so `_safe_fragments` no longer has a reachable natural trigger and is now a pure backstop, falsified synthetically. Hardening the falsifier found that this guard's own fallback re-rendered through the macro module that had just raised, so it re-raised whenever `whole` was the broken thing; fixed in the same pass. |
+| **BH-6** | Prototype pollution in the placement maps (`toString` as a mark id, `constructor` as a question key). | 1 | **Already fixed this round** as CR-13, from the code-review panel. Two panels, two lenses, the same defect independently — the strongest signal of the evening that the lenses are not redundant. |
+| **BH-7** | Bare-substring declaration suppresses the chrome. | 4 of 4 | **Already fixed** as CR-3, before the reply landed. |
+
+**One correction the panel made to this repo's own prose, adopted:** several
+comments claimed a multi-question pick POSTs a 400 unless every question is
+answered. It does not — `test_empty_submission_is_refused_with_400` refuses a
+WHOLLY EMPTY submission, and a partial answer is accepted and recorded on
+purpose. The real reason an unplaced question must still be appended is simpler
+and was being obscured: **a question that never reaches the page cannot be
+answered at all.** Fixed in `embed.js`, the browser tests and this contract.
+
+**Not adopted:** the bundle's framing called the service Flask. It is FastAPI;
+the arm noticed and declined to reason from it, which is the right handling.
+
+## Vacuity pass — final
+
+21 mutations, each drawn from an invariant's CLAIM rather than its falsifier's
+example, each run against its named test, plus an unmutated control run.
+**21/21 caught, control green.**
+
+The pass earned its place three times over and none of them was the first run:
+
+1. It reported **7/7** before the contract panel, which then showed INV-3 was
+   vacuous — because the mutation applied was the one the contract named.
+2. Re-run **against that fix**, it found the fix's own hole (an aliased
+   `import re as _r`).
+3. Re-run after the bug-hunt fixes, it reported seven **MUTATION-MISS** rows —
+   its loud-failure mode, firing correctly because the fixes had moved the code
+   out from under stale mutations — and then one genuine **VACUOUS**: the
+   sibling-mark chip test had its fixture arranged so the right answer was also
+   the first answer. Rewritten so the sibling comes first, which is the only
+   arrangement that can tell the two implementations apart.

docs/contracts/u4_derived_lifetime.contract.md

@@ -0,0 +1,606 @@
+---
+contract_version: "1.0"
+module: "booth.app (lifetime)"
+purpose: "A booth's lifetime stops being a boolean somebody remembered to press and becomes a fact derived from the booth's own state. Today there is ONE lifetime (24h from the newest mtime in the tree) and ONE escape hatch (`.forever`), and the measurement says the escape hatch is carrying the main load: 17 of 24 live booths (70%) hold the sentinel, up from the 13 of 24 (54%) counted on 2026-09-21. That is not `ephemeral with an exception`; it is two lifetimes wearing one lifetime's clothes, with the operator doing the sorting by hand. This unit adds the two facts the sweeper was missing -- a booth the operator still owes an answer to is HELD, and looking at a booth is ACTIVITY -- so the cases that were pressing `.forever` for `not yet` stop needing it, and `keep` is left meaning only what it says: this is durable."
+depends_on:
+  - "booth.marks (`hold_read` -- ADDED BY THIS UNIT, the one-read pair the lifetime rule needs; and `open_marks` -- THE openness predicate, built for this unit and saying so in its own docstring: `Open is the reading that makes U4 correct: a lifetime rule that unpinned a booth on the first radio click would sweep a review in flight.` U4 CALLS it and does not re-derive it. Also `marks_for` (lenient read, never raises) and `read_error` (strict read, total -- it catches its own `MarksCorrupt` and returns a string). Verified against booth/marks.py, not against U2's contract prose: `marks_for` is `_read_raw` + `_hydrate_safe` + sort at marks.py:514; `read_error` is `_read_raw_strict` in a try/except at marks.py:262 and has no raising path.)"
+  - "booth.items (the dotfile skip in `booth_items` at items.py -- `.viewed` is excluded from tiles, counts and zips by the EXISTING `p.name.startswith('.')` rule, exactly as `.marks.json`, `.booth.json` and `.forever` are. No new exclusion is added or needed.)"
+language: "python"
+complexity: "medium"
+estimated_loc: 130
+used_by:
+  - "booth.app.sweep_once (gains the hold check beside the keep check -- the one place reaper policy lives)"
+  - "booth.app.list_booths (the index card gains `held` and `marks_error`, so the card can say WHY it is not counting down)"
+  - "booth.app.booth_view / booth_view_file / booth_marks_page (each records a view; `/b/<n>/asks` is a 308 redirect into the last of these and so needs no call of its own)"
+  - "booth.app.booth_unkeep (release is activity -- stated, where it used to be an accident of directory mtime)"
+  - "booth/templates/index.html, booth/templates/booth.html (the lifetime line: `expires in X` / `held until answered` / `kept`)"
+touches:
+  - "booth/app.py (VIEW_MARKER, record_view, is_held; sweep_once, list_booths, booth_view, booth_view_file, booth_marks_page, booth_unkeep; the module docstring's lifetime paragraph)"
+  - "booth/templates/index.html (the ephemeral card's sub-line becomes a three-state lifetime line)"
+  - "booth/templates/booth.html (the same three-state line in the boothhead)"
+  - "booth/templates/_lifetime.html (new -- the lifetime macro, defined ONCE and called from three surfaces. Not in the first draft of this inventory: a four-state conditional repeated three times is the blurtoggle lesson, and U5 had already established the partial as the house answer.)"
+  - "booth/templates/marks.html (INV-4's third surface. A verbatim booth has no Booth-rendered header, so without this the booths most likely to be HELD -- a report that asks something -- would be the ones that never say so. Found by looking at the live service, not by the suite.)"
+  - "booth/templates/base.html (one CSS rule for the held state)"
+  - "scripts/booth (the header's `THE 24h RULE AND ITS ONE EXCEPTION` block, which states the old doctrine as the whole doctrine, and the `DO NOT unkeep and let it expire` block. The WARNING STAYS AND STAYS TRUE -- release still buys a full TTL, so unkeep-and-wait is still a delay rather than a delete. What changes is that it stops being phrased as a surprise about directory metadata and starts being phrased as the rule it now is. The paraphrase panel read the touches line as possibly meaning the advice was being retired; it is not.)"
+  - "README.md (the TTL paragraph)"
+  - "CLAUDE.md (invariant 2's dotfile list gains `.viewed`)"
+  - "tests/test_lifetime.py (new)"
+  - "tests/test_booth.py (ONE cross-reference comment. The draft said the two release-clock tests would gain an assertion that the marker is written; implementation showed they must not. `test_releasing_a_board_RESETS_its_ttl_clock` unlinks the sentinel BY HAND, not through the route, so it is a test of the mtime mechanism and asserting a route side-effect in it would be testing the wrong thing. The route behaviour is `test_releasing_a_board_RECORDS_A_VIEW` in the new file; the comment points at it. No existing assertion is touched.)"
+assumptions:
+  - "A VIEW IS RECORDED AS A DOTFILE, AND THE EXISTING AGE RULE READS IT. `.viewed` is a dotfile but NOT a `.lock` dotfile, so `_newest_mtime` already counts it (app.py:220 excludes only `.<name>.lock`). There is therefore NO new arithmetic in `booth_age_seconds`, `is_expired` or `expires_in`: `age = now - newest mtime in the tree` is unchanged, and a view is simply one more thing in the tree. One mechanism, not two. This is the same reason `.booth.json` needed no integration work in U5."
+  - "THE LOCK EXEMPTION IS WHY THIS IS SAFE. `_newest_mtime` excludes `.<name>.lock` because those are created by a READ-MODIFY-WRITE path, including one that changes nothing -- machinery, not activity. `.viewed` is the opposite: it is written only by a deliberate GET of a booth's own page. The exemption's rule (`machinery does not count, deliberate acts do`) is unchanged and this lands on the counted side of it."
+  - "RECORDING A VIEW MUST NEVER FAIL THE REQUEST. `record_view` swallows `OSError` -- a read-only mount, a booth owned by another uid, a full disk. The same posture `marks._Locked.__enter__` takes on its `os.utime` and for the same reason, stated there: `Not putting the clock back is a cost this module can absorb; not answering the request is not.` A booth that cannot record a view simply expires on its content mtime, which is today's behaviour."
+  - "HOLD IS FAIL-SAFE, WHERE READS ARE FAIL-OPEN. `marks_for` is lenient by design -- a damaged `.marks.json` reads as no marks, because a review surface that will not render is worse than one that has lost an annotation. That trade is right for a RENDER and wrong for a DELETE: the same leniency on the sweep path would wipe the booth whose judgment we had just failed to read, artifacts and all. So `is_held` treats an unreadable marks file as held. Reads lenient, deletes strict -- the same asymmetry U2 established between `marks_for` and `_Locked`, extended to the reaper. `THE REAPER` IS THE WHOLE SCOPE OF `deletes strict`, and the paraphrase panel ranked the ambiguity here first by stake: a HAND delete is never strict. `booth rm`, `POST /b/<n>/delete` and `DELETE /b/<n>` take a booth held by unreadable marks exactly as they take a kept one, which is what gives that hold -- the one nothing releases on its own -- an exit at all. Strictness is a property of the TIMER, never of the operator."
+  - "THE LIFETIME DECISION COMES FROM ONE READ, and that is a correction to this contract's first draft. The draft specified `is_held(marks_for(child), read_error(child))` -- two reads, presented as one answer. They are not: a write or a repair landing between them yields a pair that described the booth at no instant, and the losing pair is `([], None)` -- no marks and no error -- which is exactly the pair that DELETES. Hulda found it on the paraphrase round (2026-09-22) and it is the finding that changed code rather than prose. `booth.marks.hold_read(booth) -> (marks, error)` is the fix: one strict read answering both questions the lifetime rule asks, so `sweep_once` now does ONE read per booth per tick rather than two. And because `_read_raw_strict` RAISES rather than dropping an entry, a non-raising strict read returns exactly what the lenient read would -- so the index uses that same one read for its badge too, falling back to `marks_for` only on the error path, where leniency is the point."
+  - "AN OPEN PICK HOLDS; A NOTE OR A FLAG DOES NOT. `_is_open` returns False for every shape but `pick`, and False for a pick carrying `error`. That is already correct for U4 and is NOT changed here: a note is the operator's output, not an owed answer, and a pick that hydrated broken can never be answered, so holding a booth on one would be holding it forever for nothing (the CLI already spells that case as exit code 4). A PARTIALLY-answered pick IS open and DOES hold -- operator-settled 2026-09-21, and the reason `open_marks` exists rather than an `answer is None` test."
+  - "THE HOLD IS UNBOUNDED, AND THAT IS THE POINT -- BUT IT MUST BE VISIBLE. A booth with an unanswered pick is never swept, however old. This is a new way for a booth to become immortal, and it is deliberate: unanswered is unfinished. What makes it safe is not a bound, it is VISIBILITY plus TWO exits that already exist. The card and the booth header say `held until answered` in place of the countdown, so a booth that is not counting down always says why; and `booth rm` / the UI `x` delete a held booth exactly as before -- `sweep_once` is the only caller that honours a hold, precisely as it is the only caller that honours `is_kept`."
+  - "`keep` IS UNCHANGED AND KEEPS ITS LANE. `.forever` still exempts, still renders in the kept lane, still round-trips through `booth keep` / `booth unkeep` and the UI. U4 does not deprecate it, narrow it or add a reason field to it. The prediction is that its RATE falls because the `not yet` cases stop needing it -- and a prediction is falsified by measuring, not by removing the thing being measured."
+  - "THE MTIME-RESTORE RACE IN `marks._Locked.__enter__` IS EXPLICITLY CONSIDERED AND LEFT OPEN. The bug-hunt panel flagged it and it was held for U4 because closing it means changing TTL doctrine. U4's answer is that the doctrine stands: the clean fix (ignore a booth directory's own mtime whenever the booth holds anything) would close a two-syscall window that opens ONCE per booth ever, and would in exchange break every `rsync -a` populated booth -- which preserves source mtimes and so has ONLY the directory's freshness to look alive by, and which is the documented path for every host that is not nh3-dev. That is a larger hole than the one being closed. Decided, not deferred; see the OUT OF SCOPE section."
+open_questions:
+  - "Whether `booth ls` should mark held booths the way it marks kept ones with a star. Cheap, and it would need `is_held` (or a stdlib-only sibling) reachable from the CLI. Sessions already have `booth marks`, which answers the same question about their own booth, so this is convenience rather than capability. Parked, not designed."
+  - "Whether a booth held ONLY by an unreadable `.marks.json` should surface on the index as something to repair, beyond the `marks unreadable` label. It is a held booth that nothing will release, which is the one case where the unbounded hold has no natural exit. The label makes it visible; a repair affordance is a different unit."
+---
+
+# U4 — derived lifetime
+
+## The defect, stated precisely
+
+> **One lifetime (24h from last touch) and one shape (a folder), serving five
+> jobs with different lifetimes.** — `docs/design/information-architecture.md`
+
+`.forever` is the escape hatch for that mismatch, and the measurement says it is
+no longer an exception:
+
+| date | booths carrying `.forever` | rate |
+|---|---|---|
+| 2026-09-21 (IA doc) | 13 of 24 | 54% |
+| 2026-09-21 (re-count) | 14 of 25 | 56% |
+| 2026-09-22 | **17 of 24** | **70%** |
+
+Both the rate and the absolute count rose, so this is not the denominator
+shrinking as the sweeper ran. A boolean that 70% of the population sets is not
+an exception, it is the default with extra steps.
+
+The reason it gets pressed is that it is the only way to say any of these:
+
+| what the operator means | what he has to press |
+|---|---|
+| "this is a durable reference" | `.forever` |
+| "I have not answered the question yet" | `.forever` |
+| "I am still looking at this" | `.forever` |
+
+Only the first is what `keep` means. The other two are facts the service already
+holds and does not consult: **there is an open pick in `.marks.json`**, and
+**somebody just loaded the page**. U4 consults them.
+
+### The diagnosis has a live positive control
+
+Counted 2026-09-22 against `~/booth-data`. A census of the whole population, not
+a sample, and every value is a deterministic file fact (existence, mtime) — so
+one observation per booth is the measurement, not an anecdote. The population
+churns (26 -> 24 over the previous session); re-count rather than trusting these.
+
+| | |
+|---|---|
+| live booths | 24 |
+| carrying `.forever` | 17 (70%) |
+| carrying `.marks.json` at all | 4 |
+| of those, with an open pick | **4 of 4** |
+| **open pick AND `.forever`** | **3** |
+
+Three of the four booths in the fleet that are waiting on an answer have ALSO
+been pinned by hand. That is the "not yet" case, caught in the act: the operator
+pressed the durable-reference sentinel because there was no other way to say
+"do not take this, I have not answered it". U4 makes those three stop needing it.
+
+The staleness distribution says the same thing from the other side. Of the 17
+kept booths, **10 are under ONE day old** — younger than the TTL, so the
+sentinel has bought them nothing yet and was pressed pre-emptively. (An earlier
+draft of this paragraph said "12 under 1.5 days" and called that younger than
+the TTL; 1.5 days is not younger than 24 hours, and the claim only holds at the
+one-day line. Caught by the cross-frontier paraphrase panel, 2026-09-22 — the
+measurement was right and the sentence was not.) Only 4 are old enough
+(2.4-4.6 days) that `keep` is the only reason they still exist. A
+sentinel pressed on a booth that was in no danger is not a durability decision;
+it is "not yet", written in the only vocabulary available.
+
+⚠ The hold's live blast radius is SMALL today — 4 booths have marks at all. The
+17-to-something prediction therefore rests on both halves of this unit, and on
+the sentinel becoming unnecessary rather than becoming forbidden. If the rate
+does not move, the honest readings are: the diagnosis was wrong, OR the habit
+outlived the need, and the fortnight re-count cannot tell those apart on its
+own. The three open-pick-plus-`.forever` booths are the ones to watch, because
+for them the mechanism is now unambiguous.
+
+## The record
+
+A booth is in exactly one lifetime state, decided in this order:
+
+```
+KEPT     .forever present            never swept          (unchanged)
+HELD     an open pick, or a          never swept while    (new)
+         .marks.json we cannot read  that holds
+EPHEMERAL otherwise                  swept when
+                                     age > ttl            (unchanged)
+```
+
+`age` is unchanged: `now - _newest_mtime(booth)`, the newest mtime in the tree
+excluding `.<name>.lock`. **Viewing is folded in through that existing rule**,
+not beside it — a view writes `.viewed`, which is a dotfile and not a lock
+dotfile, so the age function already counts it. There is no new arithmetic.
+
+### What counts as a view
+
+One line, because CLAUDE.md invariant 6's test applies to rules as well as
+orders: **a deliberately-requested response FROM a booth's own page route is a
+view; a machine read, an asset fetch, and a request that does not resolve are
+not.**
+
+Three words in that rule are load-bearing and the first draft said "HTML page",
+which was wrong twice. `?download=1` is a zip served by the booth-page route and
+IS a view — the operator asking for the whole booth is as deliberate as looking
+at it. And a `/view?f=<missing>` that 404s is NOT one: the route matters, but so
+does whether anything was served, or a crawler walking dead zoom URLs holds a
+booth open forever. `record_view` therefore sits below the zoom route's file
+validation and above the booth route's verbatim/zip fork.
+
+| route | view? | why |
+|---|---|---|
+| `GET /b/<n>/` | **yes** | the booth page — gallery, verbatim report, or `?download=1` zip |
+| `GET /b/<n>/view?f=…` | **yes** | the zoom / doc page; a bookmarked zoom URL is somebody looking |
+| `GET /b/<n>/marks` | **yes** | the standalone judgment page — for a verbatim booth this IS the booth page |
+| `GET /b/<n>/marks.json` | no | a session polling. An agent must not be able to hold its own booth open |
+| `GET /b/<n>/<file>` | no | issued BY the page. A hotlinked image would otherwise keep a booth alive |
+| `GET /` | no | the IA's rule: "deliberate act, so it cannot be triggered by browsing the index" |
+| `GET /healthz` | no | a monitor is not a viewer |
+
+⚠ Named rather than hidden: `scripts/layout-probe.py` sweeps every booth page,
+so running it resets every booth's clock. That is the correct reading of the
+rule (it is a GET of every booth page), it is recoverable (one extra TTL), and
+it is a dev tool. A note goes in the probe.
+
+⚠ A browser that speculatively prefetches a hovered link records a view the
+operator did not quite take. Accepted: the failure mode is a booth living one
+extra day because he nearly opened it, and the alternative is sniffing
+`Sec-Fetch-*` headers, which is a fragile rule pretending to be a crisp one.
+
+**Checked, because it would have been silent:** nothing in the fleet polls a
+booth *page*. Homepage's `siteMonitor` for the Booth is
+`http://10.100.10.50:8090/healthz`, which is on the not-a-view list; there is no
+cron entry and no systemd timer touching `/b/…`. Had Homepage been pointed at a
+booth URL instead, every booth would have become immortal on deploy and nothing
+would have reported it.
+
+### Release is activity, on purpose
+
+Removing `.forever` bumps the booth directory's mtime, so a released board
+survives another full TTL. Today that is an **accident** of directory metadata
+that `app.py` documents as "not intuitive" and `scripts/booth` warns against.
+
+U4 does not change the behaviour and does not retire the test that pins it. It
+changes the behaviour's *reason*: `booth_unkeep` calls `record_view`, so a
+released board gets one full TTL because **releasing a board is somebody
+touching it**, which is a rule, and no longer because of which syscall happened
+to write a directory entry, which is not.
+
+The existing tests (`test_releasing_a_board_RESETS_its_ttl_clock`,
+`test_released_board_is_sweepable_once_it_ages_again`) are untouched, and that
+is a correction to this contract's first draft, which said they would each gain
+an assertion that the marker is present. They must not: the first one unlinks
+the sentinel **by hand**, not through the route, so it is a test of the mtime
+mechanism and a route side-effect does not belong in it. The route behaviour
+gets its own test in the new file, and the old test gains a comment pointing at
+it.
+
+**The marker's mtime must be NOW**, which `Path.touch()` gives and which the
+contract's first draft left unsaid. An implementation that wrote the file with
+any older timestamp would satisfy "the marker is there" while the extra TTL
+still came from the directory-mtime accident this section exists to replace —
+the new reason would be decoration over the old mechanism. Flagged by the
+paraphrase panel, 2026-09-22.
+
+## Signatures
+
+```python
+# booth/app.py
+
+VIEW_MARKER = ".viewed"
+"""Records the last deliberate look at a booth. A dotfile, so `booth_items`
+skips it and it costs nothing in counts, galleries or zips — and NOT a `.lock`
+dotfile, so `_newest_mtime` counts it and the existing age rule picks up the
+view with no new arithmetic."""
+
+
+def record_view(booth: Path) -> None:
+    """Note that somebody deliberately looked at this booth.
+
+    Touches VIEW_MARKER; `_newest_mtime` does the rest. NEVER raises: a
+    read-only mount, a booth we do not own or a full disk cost the timestamp,
+    not the page. A booth whose view cannot be recorded simply ages on its
+    content mtime, which is today's behaviour for every booth.
+    """
+
+
+def is_held(marks: Sequence[Mark], error: str | None) -> bool:
+    """True if this booth still owes the operator an answer and must not be swept.
+
+    PURE — it takes the result of a read and does none of its own, so the index
+    card and the sweeper cannot answer differently about the same booth. That
+    is U1's rule (one resolver, every surface reads the record) applied to
+    lifetime.
+
+    FAIL-SAFE on `error`. `marks_for` is lenient because a review page that
+    will not render is worse than one missing an annotation; the same leniency
+    on the DELETE path would wipe the booth whose judgment we had just failed
+    to read. Reads lenient, deletes strict.
+
+    Openness itself is `open_marks` and nothing else (U2 INV-2).
+    """
+    return error is not None or bool(open_marks(marks))
+```
+
+`is_expired` is **unchanged** and stays a pure age question — the existing
+separation ("expiry arithmetic and reaper policy are kept apart so they cannot
+drift into each other") is the reason `is_kept` is not consulted there either.
+`sweep_once` remains the only caller that honours a pin, and now honours two.
+
+```python
+# booth/marks.py — stdlib only, like the rest of that module
+
+def hold_read(booth: Path) -> tuple[list[Mark], str | None]:
+    """ONE read of `.marks.json`, answering BOTH questions the lifetime rule
+    asks: what is still open, and whether the file could be read at all.
+
+    Two calls would read the file twice, and two reads of one file are not one
+    read of one state — the pair that loses the race is `([], None)`, which is
+    the pair that deletes.
+
+    On a clean file the marks are what `marks_for` would return, because
+    `_read_raw_strict` raises rather than dropping an entry. So one read serves
+    the badge too, and the lenient reader comes back only on the error path.
+    """
+```
+
+```python
+def sweep_once(data_dir, ttl_seconds, now=None) -> list[str]:
+    ...
+        if is_kept(child):
+            continue
+        if is_held(*hold_read(child)):                     # NEW — ONE read
+            continue
+        if is_expired(child, ttl_seconds, now):
+            shutil.rmtree(child)
+```
+
+```python
+def list_booths(data_dir, ttl_seconds, now=None) -> list[dict]:
+    ...
+    marks, marks_error = hold_read(child)                  # NEW — one read, both facts
+    if marks_error is not None:
+        marks = marks_for(child)                           # lenient, for the panel
+    booths.append({
+        ...
+        "marks_error": marks_error,                        # NEW — the card says so
+        "held": is_held(marks, marks_error),               # NEW — the same predicate
+    })
+```
+
+## What renders
+
+The lifetime line, on the ephemeral index card and in the booth header. Three
+states, one of which is new:
+
+| state | line | why |
+|---|---|---|
+| ephemeral | `12 items · expires in 3h 20m` | unchanged |
+| held, open pick | `12 items · held until answered` | says what holds it AND what releases it |
+| held, unreadable | `12 items · held · marks unreadable` | the one hold nothing will release on its own |
+| kept | `12 items · kept` | unchanged, kept lane |
+
+**The hold REPLACES the countdown at every age, not only once the booth is
+old.** A held booth that is four hours old shows `held until answered`, not
+`expires in 20h`. `expires_in` is still computed and still correct (INV-1);
+it is simply not what the surface says, because a number counting down to a
+deletion that will not happen is the silent-stopped-clock failure in its other
+costume — the screen announcing an expiry the sweeper will never carry out.
+Flagged as readable-two-ways by the paraphrase panel, 2026-09-22; settled here.
+
+A booth that is not counting down **always says why**. That is the whole safety
+argument for an unbounded hold: `.forever` was at least visible as a lane; an
+invisible rule that silently stops the clock would be strictly worse than the
+boolean it replaces.
+
+**Three surfaces, not two**, and the third was found by looking at the live
+service rather than by the suite. A verbatim booth's own `index.html` is served
+untouched by design, so it has no Booth-rendered header for the line to live in
+— and a report that ASKS the operator something is the archetype of a held
+booth. `GET /b/<n>/marks` is the only other page whose chrome the Booth owns, so
+the line goes there too. Without it, the booths most likely to be held would be
+exactly the ones that never said they were. (U3 is the unit that gives a
+verbatim booth real chrome; until then, this is the honest coverage.)
+
+Kept beats held in the display, because a kept booth is in the kept lane and is
+exempt either way — showing two reasons for one exemption is the
+two-representations-of-one-state trap `flag_id`'s docstring names.
+
+**An unreadable marks file is the exception, and it rides along even on a kept
+board**: `kept · marks unreadable`. Damaged judgment is not a second exemption,
+it is a thing somebody has to go and fix, and the kept lane holds the durable
+boards — the ones where losing the operator's marks costs most. A kept card that
+said only `kept` would hide the single case that needs a human. The card's
+`held` and `marks_error` are therefore RAW FACTS, true regardless of keep, and
+only the display has a precedence. The paraphrase panel found the two readings
+of "exactly one lifetime state" that this settles.
+
+## Scope — the blast-radius pass
+
+`graphify explain` on `sweep_once`, `is_kept`, `list_booths`, `_newest_mtime`,
+`booth_age_seconds`, `open_marks`, `KEEP_MARKER`, cross-checked with grep.
+
+Graphify reported the call structure and, as expected, **missed both route
+callers of `list_booths`** (`index()` and `healthz()`, now at app.py:761 and :774) —
+they are function-local inside `create_app`, which is the known AST blind spot.
+Grep caught them. Neither tool alone was sufficient; this is the third unit in
+a row where that has been true.
+
+**Production, 7 files:** `booth/app.py`, `booth/marks.py` (`hold_read`, added),
+`booth/templates/_lifetime.html` (new), `booth/templates/index.html`,
+`booth/templates/booth.html`, `booth/templates/marks.html`,
+`booth/templates/base.html`.
+**Docs/CLI, 4 files:** `scripts/booth`, `scripts/layout-probe.py`, `README.md`,
+`CLAUDE.md`.
+**Tests, 2 files:** `tests/test_lifetime.py` (new), `tests/test_booth.py`.
+
+⚠ This census said "Production, 4 files" in the first draft and omitted
+`_lifetime.html`, `marks.html`, `marks.py` and `layout-probe.py` — three of
+which the body text elsewhere required, which is the contradiction both
+Gróa and Hulda flagged independently. An inventory that disagrees with the
+prose next to it is worse than no inventory: it reads as a closed set.
+
+Not touched, and checked rather than assumed: `booth/items.py`,
+`booth/manifest.py`, `booth/links.py`, `booth/asks.py`, `booth/inline.py`.
+
+## The three cross-frontier panels, and what they changed
+
+All three ran on 2026-09-22 and all three earned their place — and each found
+a class the other two could not. Triaged per the cross-frontier discipline
+rather than adopted.
+
+**Paraphrase panel** (`01M34VX0SH23Y3VC92E7GM4S70`, four arms). Seven flags.
+Five folded into the prose above: the hold replacing the countdown at every
+age, `deletes strict` scoping to the reaper alone, the zip and the 404 in the
+view rule, the marker's mtime, and the CLI warning staying true. Two changed
+more than wording:
+
+- **Hulda — the two reads are not one state.** The only finding on this round
+  that changed CODE. See the `hold_read` assumption in the frontmatter.
+- **Gróa and Hulda, independently — the blast-radius census contradicted the
+  prose beside it.** It named four production files while the body required
+  three more. An inventory that disagrees with its own document is worse than
+  none, because it reads as a closed set.
+
+Hulda also caught a number: this contract claimed 12 kept booths were "under
+1.5 days old — younger than the TTL". One and a half days is not younger than
+twenty-four hours. The measurement was right, the sentence was not, and it is
+the one place the diagnosis overstated itself.
+
+**Code-vs-contract panel** (`01M34WAFJC3RTERFYBBZJN1SVG`, four arms). **All
+four arms found the same drift** — the strongest signal either panel produced
+on this unit. The booth header's sub-line forks on `{% if board %}`, and the
+lifetime macro sat only in the `{% else %}`, so a booth carrying `links.md`
+rendered a link count and nothing at all about its lifetime. INV-4 says the
+templates have no path that renders neither; that was a path, reachable by the
+release button or by a hand-made board.
+
+Regin and Kimi recommended amending INV-4 to carve the board header out, on the
+grounds that board-header layout belongs to U7. **Declined; the code is fixed
+instead.** Cutting an invariant down to fit an implementation gap is the wrong
+direction when the fix is one template edit, and U7 owns navigation and section
+layout — not whether a header states a lifetime. Gróa's "fix it" was right.
+
+The same panel showed that **most of the INV falsifier tests did not
+discriminate**, which is the more useful half of the round. The header test
+never rendered a board. The kept-beats-held test only rendered the index, where
+kept cards took a hardcoded string and never reached the macro at all. The
+INV-5 test called `record_view` directly instead of GETting the routes the
+invariant is about. The INV-7 tests asserted the marker's absence rather than
+the age, so a handler writing any other non-dot file would have passed. INV-6's
+had no doomed sibling, so "spare everything" would have passed. Each is now
+written to fail under the change that defeats it, and the board-header pair was
+verified RED against the pre-fix template rather than assumed.
+
+**Bug-hunt panel** (`01M34Y2R0RAJRSN36Q8K4KAB36`, four arms). The round that
+changed the most code, and the one that found a class the other two could not
+see by construction: **a read that FAILED still resolving to "no hold", and
+therefore to a delete.** That is the invariant this unit declared to the panel,
+and the panel found **four independent paths through it. No single arm found
+all four.**
+
+1. **An entry-level hydration error lost its hold.** `.marks.json` parses, one
+   mark fails normalization, `_hydrate_safe` returns a `Mark` carrying `error`,
+   and `_is_open` returns False for an errored pick — on purpose, because a
+   broken pick can never be answered. So the booth read as not-held and swept,
+   while the panel beside it rendered the broken mark in full. The fail-safe was
+   built for FILE-level damage and missed ENTRY-level. This is the strongest
+   finding of all three rounds.
+2. **A present-but-blank `.marks.json` swept.** `_read_raw_strict` early-returns
+   for whitespace-only content — right for the write path it was written for,
+   wrong for the delete path. Our writer never produces a blank marks document,
+   so a blank one that exists is something that went wrong.
+3. **`_newest_mtime` returned 0.0 when the booth's own stat failed**, making it
+   maximally ancient and therefore the FIRST thing the sweeper takes. Pre-dates
+   U4; U4 is what turned the age read into a life-or-death read.
+4. **`is_kept` collapsed a stat failure into not-kept.** `Path.exists()` maps
+   ELOOP and EACCES to False, so a kept booth whose sentinel could not be
+   stat'd became sweepable.
+
+**`is_held` is gone; `hold_reason` replaced it.** A boolean plus a separate
+error string is two representations of one state, and Regin independently
+flagged that the display could not distinguish the two holds. One function now
+returns the REASON — `"open"`, `"unreadable"`, or None — and every surface reads
+it off the same value the sweeper acts on. That closes findings 1 and Regin's
+together, which is why it is a rewrite rather than an extra clause.
+
+**Convergent, 3-of-4: `record_view` followed a planted symlink.** `Path.touch()`
+follows an existing link, so a booth carrying `.viewed -> /anywhere` turned every
+page view into an mtime write at an arbitrary path under the service uid — and
+any fleet session can write into a booth, because making a folder is the whole
+API. Now an `O_NOFOLLOW` create plus `os.utime(fd)`, so a planted link raises
+ELOOP into the existing swallow and view-recording quietly stops for that booth.
+The `utime` is also what makes the marker read as NOW, which this contract
+already required and `O_CREAT` alone does not do.
+
+**Two more the panel found in code this unit touched:**
+
+- **`?f=.marks.lock` held a booth open.** The zoom route recorded a view for any
+  path that stats inside the booth, including a lock file the service created
+  itself. `record_view` now sits below `find_item` and fires only for a real
+  item — which also makes the comment beside it true, where before it claimed
+  more than the code did.
+- **Releasing an ALREADY-released booth refreshed its TTL forever.** The
+  unconditional `record_view` on `unkeep` contradicted that route's own no-op
+  promise and diverged from the CLI, which removes the sentinel without
+  recording anything. Now gated on something actually having been released. The
+  same edit fixes a pre-existing 500: a `.forever` that is a DIRECTORY raised
+  `IsADirectoryError` straight through the route, which made the card's release
+  button permanently dead for that booth.
+
+**Also fixed: a docstring this unit's own fix made stale.** `sweep_once` still
+claimed "one lenient read plus one strict read" after `hold_read` reduced it to
+one. Kimi's framing is the right reason to care — a maintainer "optimizes" back
+to two calls on the comment's authority, and rebuilds the seam the function
+exists to kill.
+
+**Re-declared as parked, not adopted:** Regin distinguished a stale-DECISION
+window (hold checked, then rmtree) from the torn-FILE race already parked at
+`park/booth-sweeper-rename-then-delete-to-close-the`. The distinction is real
+and the fix is the same rename-then-delete, so it parks with its sibling.
+
+**Five pre-existing defects the panel surfaced in touched files** — a booth name
+reaching a JS string context, an unguarded `links.md` read, an index sort with
+no tie-breaker, `marks.json` reporting damage as empty success — are fixed in
+their own commit rather than smuggled into this unit's. See that commit.
+
+⚠ **The capture tooling failed silently and the panel caught it, not us.** The
+`files/` tree shipped to the arms was EMPTY: the snapshot loop iterated `for f
+in $IN` over a multi-line variable, and **zsh does not word-split unquoted
+parameter expansions** the way bash does, so it ran once against a path that was
+the whole list. jekyll recovered by re-applying the bundled diff to HEAD and
+verified every file byte-identical, so the round is sound — but the failure mode
+is the dangerous one: an empty bundle reads exactly like a clean result.
+
+## Seam review — against the real module surface
+
+Checked against `booth/marks.py` itself, not against U2's contract prose.
+
+| borrowed | real surface | verdict |
+|---|---|---|
+| `open_marks(marks)` | `marks.py:542`, takes `Sequence[Mark]`, returns `list[Mark]` | matches |
+| `marks_for(booth)` | `marks.py:514`, `_read_raw` + `_hydrate_safe` + sort; total | matches |
+| `read_error(booth)` | `marks.py:262`, returns `str \| None`, catches its own `MarksCorrupt` | matches — **and it is total**, which `is_held`'s fail-safe branch depends on |
+| `_is_open` semantics | `marks.py:525`: `pick` only, `error is None`, partial counts open | matches the assumption above |
+| `_newest_mtime` lock rule | `app.py:220`: skips `p.name.startswith(".") and p.name.endswith(".lock")` | `.viewed` is counted — confirmed at the source, not inferred |
+| `Mark` import in app.py | app.py:145-160 imports `open_marks`, `marks_for`, `marks_for_target`, `as_dict` — **not `Mark`** | `is_held`'s annotation needs `Mark` added to that import list |
+| `read_error` import in app.py | **not imported either** — U2 left it to the CLI, which is its only caller today | must be added to the same block; U4 is its first in-service consumer |
+| `zip_booth` dotfile skip | `app.py:445`ff: `p.is_file() and not p.name.startswith(".")` | `.viewed` never reaches a zip — confirmed, not inferred from `booth_items` |
+| `booth_items` dotfile skip | `items.py:182`: `not p.is_file() or p.name.startswith(".")` | `.viewed` is not an item |
+| `GET /b/<n>/asks` | `app.py:1105`, a **308 redirect** to `/marks`, not its own render | records a view through the `/marks` handler. No separate call, and adding one would double-count |
+| route concurrency | `booth_view`, `booth_view_file`, `booth_marks_page` are all `def`, not `async def` | FastAPI runs them in a threadpool, so `record_view`'s write cannot block the event loop |
+
+Three rows of that table are the kind of thing only this pass finds: the cold
+panel reads one contract, and a signature that is fine in isolation says nothing
+about whether the name it needs is in scope at the call site.
+
+**SR-1 — why `read_error` is safe to call per booth per index load, which the
+signatures alone do not say.** `_read_raw_strict` checks `S_ISREG` *before* it
+calls `read_text` (marks.py:236). That ordering is the v0.2.2 fix: `st_size` is
+0 for a FIFO and 0 for a symlink to `/dev/zero`, so a size cap alone lets both
+through and `read_text` then either blocks with no EOF or allocates until the
+kernel intervenes — across every booth, on `GET /`, which is a service-wide
+hang rather than one bad card. U4's decision to spend a second read on the hot
+path depends on that guard already being there. It is; checked at the source.
+
+## Out of scope
+
+- **A bound on the hold.** An abandoned pick holds its booth forever. Detecting
+  "abandoned" needs state the Booth does not have (is any session still
+  polling?), and the honest alternative — an arbitrary N-day cap — trades a
+  visible immortal booth for a silent deletion of an open question. Visibility
+  plus `booth rm` is the answer for v1.
+- **A reason string on `.forever`.** "keep survives as an explicit, reasoned
+  pin" is read here as *a pin the operator reasoned about*, not *a pin carrying
+  a recorded reason*. A `why` on keep does not close the measured defect — the
+  70% is people using keep for things that are not keep, and this unit gives
+  those things their own mechanism. Parked per the anti-creep gate.
+- **A third index lane for held booths.** A booth waiting on the operator is the
+  most actionable thing on the index, and it already carries the `? N open`
+  badge. Lane structure and ordering are U7's, and adding a lane here would set
+  an ordering rule that U7 then has to live with.
+- **Closing the `marks._Locked.__enter__` mtime-restore race.** See the
+  assumption above: the clean fix costs every `rsync -a` populated booth. The
+  comment there stays, and stays accurate.
+- **`booth ls` marking held booths.** Open question, parked.
+- **Closing the view-during-sweep race, which U4 WIDENS.** `sweep_once` calls
+  `shutil.rmtree` without holding anything, so a write landing inside that call
+  can make it raise partway and leave a stump directory. The race is
+  pre-existing — every write route has always had it — but U4 widens it,
+  because `record_view` fires on every booth-page GET and the case that
+  collides is precisely "the first look at a booth that has been silent for 24
+  hours", which is the state the sweeper acts on.
+
+  The fix is known and small: `os.rename` the booth to `.sweeping-<name>` first
+  (atomic, and a dotfolder the scan already skips), then `rmtree` the renamed
+  path, plus a cleanup of leftovers at the top of each tick for the
+  crash-between-the-two case. It is NOT done here, per the anti-creep gate:
+  both "in" and "park" are defensible, so it parks. The arithmetic is that the
+  collision needs a GET inside a ~10 ms `rmtree` on a booth nobody has opened in
+  a day, the sweeper ticks every 15 minutes, and the consequence is a stump that
+  survives one more TTL — against which a sweeper rewrite is not a v1-path
+  trade. Named here so it is a decision and not an oversight, and parked on
+  the henge at `park/booth-sweeper-rename-then-delete-to-close-the` (id 83)
+  so it has a home rather than only a paragraph.
+
+## Invariants
+
+**INV-1 — Age arithmetic is unchanged.** `booth_age_seconds`, `is_expired` and
+the `expires_in` values on both surfaces are computed exactly as before. A view
+enters through `_newest_mtime` as a file in the tree, not as a term in a new
+formula. *Falsifiable:* a booth with a `.viewed` and a booth with any other
+non-lock dotfile of the same mtime report the same age.
+
+**INV-2 — `sweep_once` is the only caller that honours a hold.** `is_expired`
+stays a pure age question; `booth rm`, `POST /b/<n>/delete` and
+`DELETE /b/<n>` delete a held booth exactly as they delete a kept one.
+*Falsifiable:* a held booth is still reported expired by `is_expired` and is
+still deleted by the delete routes.
+
+**INV-3 — One predicate, ONE READ, one answer.** The index card's `held`, the
+booth header's, the marks page's and the sweeper's exemption all come from the
+same pure `is_held`, and each call's two inputs come from a SINGLE read of
+`.marks.json` via `hold_read` — never from two reads stitched together, which
+is a pair that described the booth at no instant. *Falsifiable:* for any booth, what
+`list_booths` reports as `held` and what `sweep_once` refuses to take agree —
+tested directly rather than by inspection, because that is the falsifiable form.
+(`open_marks` is still called directly for the `N open` COUNT. A count is not a
+lifetime decision, and the first draft of this invariant forbade it by accident
+— the rule is that no *exemption* and no *held label* is derived except through
+`is_held`.)
+
+**INV-4 — A booth that is not counting down says why.** Every non-kept booth
+renders either a countdown or a named hold on **every surface whose chrome the
+Booth owns**: the index card, the booth header, and the marks page (which is
+the only one of the three a verbatim booth has). *Falsifiable:* the templates
+have no path that renders neither, and the one line is a single macro rather
+than three conditionals that can drift.
+
+**INV-5 — Recording a view cannot fail a request.** `record_view` swallows
+`OSError`. *Falsifiable:* a booth whose directory is read-only still returns 200
+for its page, its zoom page and its marks page.
+
+**INV-6 — An unreadable `.marks.json` holds its booth.** The reaper never
+deletes judgment it could not read. *Falsifiable:* a booth with a corrupt
+`.marks.json`, aged past the TTL, survives `sweep_once`.
+
+**INV-7 — Machine reads do not hold a booth open.** `GET /b/<n>/marks.json` and
+`GET /b/<n>/<file>` do not write `VIEW_MARKER`. *Falsifiable:* polling either,
+repeatedly, leaves the booth's age untouched.

docs/contracts/u5_booth_manifest.contract.md

@@ -0,0 +1,405 @@
+---
+contract_version: "1.0"
+module: "booth.manifest"
+purpose: "A booth that says what it IS and who posted it. Today the index card shows a name, an item count and a countdown -- nothing about provenance or purpose -- so an agent that wants the operator to look at something has no way to make the booth say so, and posts a URL to the link board instead. That is job 5 (`Announce`), the job nobody named, and its absence is the measured cause of 145 dead link rows (69% of the board pointing at booths that no longer exist). This unit gives job 5 a home: each booth carries `.booth.json` -- `{handle, title, why, created}`, written by the CLI from `$ALTHING_HANDLE` -- and the index card and the booth page header render it. Enforcing the link rule WITHOUT giving job 5 a home first just makes it homeless; this is the home."
+depends_on:
+  - "booth.items (the dotfile skip in `booth_items` -- `.booth.json` is excluded from tiles, counts and zips by the EXISTING `p.name.startswith('.')` rule at items.py:182, exactly as `.marks.json` is. No new exclusion rule is added or needed. Verified, not assumed: `test_a_manifest_is_not_an_item` asserts it.)"
+  - "booth.marks (the `_write_raw` shape only -- temp file + os.replace, per CLAUDE.md invariant 5. Copied as a pattern, NOT imported: manifest.py must not depend on marks.py, because the CLI imports each module on its own.)"
+language: "python"
+complexity: "low"
+estimated_loc: 150
+confidence: 0.85
+used_by:
+  - "booth.app.list_booths (the index card gains `manifest` -- one file read per booth, alongside the `marks_for` read already there)"
+  - "booth.app.booth_view (the booth page header gains the same provenance line; a booth URL handed to the operator lands HERE, not on the index, and job 5 is literally 'operator, look at this')"
+  - "booth.app.upload (a pickup booth announces itself as the Booth's own)"
+  - "scripts/booth (`new` and `add` gain `--why` / `--title`; `link` announces the standing board)"
+touches:
+  - "booth/manifest.py (new -- the record, the write, the lenient read)"
+  - "booth/app.py (list_booths gains one key; booth_view gains one key; the /upload path writes a manifest. It also adds MANIFEST_FILE to the `used` dedupe set -- CONSISTENCY, not a fix: SR-1 established the collision is unreachable because `safe_upload_name` strips leading dots, which is equally true of the `UPLOAD_MARKER` entry that has sat in that set since before this unit.)"
+  - "booth/templates/_provenance.html (new -- the provenance macro, defined ONCE and called from both index lanes and the booth header. Not in the first draft of this inventory: the implementation added the partial rather than repeating the four-state conditional three times, which is SR-6 plus the blurtoggle lesson, and the inventory lagged the decision.)"
+  - "booth/templates/index.html (the provenance line on both lanes' cards -- kept AND ephemeral, or the kept lane silently keeps the old defect)"
+  - "booth/templates/booth.html (the provenance line in the boothhead, and the h1 renders `title` with the directory name beside it)"
+  - "booth/templates/base.html (the .prov-* CSS)"
+  - "scripts/booth (`new` / `add` flag parse; `link` board announcement; usage string; the header doc block)"
+  - "tests/test_manifest.py (new)"
+  - "tests/test_marks.py (test_stdlib_only's parametrize list gains `manifest`)"
+assumptions:
+  - "THE MANIFEST IS A DOTFILE, and that is the whole integration story. `booth_items` skips `name.startswith('.')` (items.py:182), `zip_booth` skips it (app.py:351), and the legacy ask scan skips it (marks.py:656). So `.booth.json` costs nothing in item counts, galleries, zips or migration, and needs no new exclusion anywhere. This is the same reason `.marks.json` needed none. Settled -- do not re-derive it."
+  - "WRITING A MANIFEST IS ACTIVITY. `.booth.json` is a dotfile but NOT a `.lock` dotfile, so `_newest_mtime` counts it (app.py:192 excludes only `.<name>.lock`). Creating or re-announcing a booth resets its TTL, which is correct: both are somebody touching it. The lock exemption exists for machinery that a READ path creates; this is a deliberate write."
+  - "THE READ IS LENIENT AND THE FAILURE IS VISIBLE. `list_booths` reads every booth on every index load, so a manifest that cannot be parsed must never raise -- that is the v0.2.2 lesson, learned when a poisoned `.marks.json` returned 500 for `/` and `/healthz` across all 25 booths. `read_manifest` returns None for absent and a `Manifest` carrying `error` for damaged, and the card distinguishes them (`unannounced` vs `unreadable`). Silently treating damaged as absent would hide the one case somebody has to fix."
+  - "THE WRITE IS ATOMIC (CLAUDE.md invariant 5, NOT this unit's INV-5). Temp file + os.replace onto a name no other writer derives, because the CLI writes it in one process while the browser reads it in another -- and because two `booth add` calls on one booth would otherwise share a scratch name, which the atomic-write promise says nothing about: it promises readers never see a partial file, not that writers never race. The pattern is copied from `marks._write_raw` rather than imported: `scripts/booth` imports each module directly under the system python3, and a cross-import between two stdlib-only modules is a second way for INV-1 to break."
+  - "`booth/manifest.py` IS STDLIB-ONLY and joins the CLAUDE.md invariant 1 list. `scripts/booth` imports it through a `python3 -c` heredoc with no venv, exactly as it imports `marks`, `asks` and `links`. `test_stdlib_only` is parametrized and gains `manifest`; that test is the only thing standing between a casual third-party import and `booth new` breaking on every fleet host."
+  - "A MISSING MANIFEST IS NORMAL, NOT AN ERROR. All 26 live booths have none, and `rsync -a ./out/ nh3-dev:booth-data/my-run/` -- the documented path for every host that is not nh3-dev -- never runs the CLI at all, so unannounced booths keep arriving after this lands. The card marks them quietly and nothing refuses to render, expire, zip or sweep."
+  - "THE BOOTH ANNOUNCES ITS OWN BOOTHS rather than exempting them. A pickup booth and the standing link board are created BY the service, so they are written with `handle: booth` -- which is true, not manufactured. The alternative was a pile of exemptions from the unannounced marker; this way there is one rule (a booth with no manifest is unannounced) and no special cases. `handle` therefore names an agent handle OR the service, and the field's docstring says so."
+  - "NOTHING NEW IS ORDERED, so CLAUDE.md invariant 6 (every ordered collection has a stated, deterministic rule) does not bind here -- there is no new collection for it to bind to. The manifest is one flat record per booth. The index keeps its stated rule -- kept lane first, then ephemeral newest-first by `_newest_mtime` -- and U5 does NOT add a second ordering keyed on `created` (operator, 2026-09-22). A what-landed feed ordered by announcement time is a genuinely different surface: it needs its own stated rule, it competes with the existing order for what 'the third one' means, and it has nothing to sort the 26 manifest-less booths by. Parked for v1.1."
+open_questions:
+  - "Whether `why` should also reach the zip manifest or a `booth ls` column. Both are one-liners over the same record and neither is on the v1 path; deferred rather than designed."
+---
+
+# U5 — self-announcing booths
+
+## The defect, stated precisely
+
+The index card is the only thing an agent can put in front of the operator, and
+it carries no information the agent chose. Name, item count, countdown, a
+thumbnail. Everything about *why this exists* has to travel some other way.
+
+So it travelled some other way. `booth link` exists because a session with
+something to show had no way to make the booth itself say "look at this", and
+the link board absorbed job 5 until **145 of its 210 rows (69%) pointed at
+booths that had already been swept**. The rot is not a link-board bug. The board
+was doing a job it was never shaped for, because the shaped thing did not exist.
+
+The lesson the measurement carries, and the reason this unit comes before any
+link-board enforcement: **enforcing the link rule without giving job 5 a home
+just makes it homeless.**
+
+## The record
+
+```python
+@dataclass(frozen=True)
+class Manifest:
+    handle: str            # an althing handle, or "booth" for one the service made
+    title: str             # display name; falls back to the directory name
+    why: str               # ONE line: what the operator is looking at and why
+    created: str           # ISO-8601 with offset, from the FIRST announcement
+    error: str | None = None   # a read-time verdict; never stored
+```
+
+`.booth.json` on disk is the same four fields, no `error`.
+
+**Every field on an error-carrying record has a stated value**, because the
+templates render the record and a careless fill would re-raise the outage in
+the renderer: `handle` and `why` and `created` are `""`, `title` is the
+normalized directory name, and `error` says which of the six refusals fired.
+`created` being `""` is what makes `write_manifest` treat a damaged prior as
+having no stamp to preserve (INV-3).
+
+Caps, all applied at the write and again at the read: `handle` 64, `title` 120,
+`why` 200, `created` 64. Each is a **display budget**, not a storage limit —
+they exist because these strings land in a card's sub-line.
+
+## Signatures
+
+```python
+MANIFEST_FILE = ".booth.json"
+HANDLE_MAX, TITLE_MAX, WHY_MAX = 64, 120, 200
+
+
+MANIFEST_MAX_BYTES = 64 * 1024
+QUARANTINE_FILE = ".booth.json.broken"
+
+
+def read_manifest(booth: Path) -> Manifest | None:
+    """This booth's announcement, or None if it never made one.
+
+    LENIENT, and never raises. `list_booths` calls this once per booth on every
+    index page load, so a damaged file must cost that booth's provenance and
+    nothing else — the same posture `marks_for` takes, for the reason v0.2.2
+    made expensive: a read that can raise, called in a loop over every booth,
+    is a service-wide outage wearing a single-booth bug's clothes.
+
+    "NEVER RAISES" IS BOUNDED, NOT MERELY CAUGHT. An earlier draft of this
+    contract named a 4 GB file as a tested case and constrained only the RETURN
+    — which is letter-compliant and purpose-defeating: reading four gigabytes
+    per booth per index load recreates the same outage in slow motion. The size
+    is checked by `stat` BEFORE the bytes are touched, and the two exception
+    classes that are neither `OSError` nor `ValueError` — `MemoryError` from a
+    huge document, `RecursionError` from a deeply nested one — are caught as
+    well, so that raising the bound one day cannot quietly re-open the hole.
+
+    REGULAR-FILE FIRST, THEN SIZE — and the order is the whole point. `st_size`
+    is 0 for a FIFO and 0 for a symlink to `/dev/zero`, so both sail under any
+    byte cap and then the read either blocks forever with no EOF or allocates
+    until the kernel intervenes. The bound is what made this reachable: a cap
+    that trusts `st_size` inherits everything `st_size` does not mean. One such
+    file stalls every `GET /` and `/healthz`, with no error and no recovery
+    short of a restart.
+
+    Absent -> None. Present but too large, unreadable, unparseable, not an
+    object, or missing `handle` -> a Manifest carrying `error`, so the card can
+    say `unreadable` rather than quietly showing the same thing as a booth that
+    never announced.
+    """
+
+
+def write_manifest(booth: Path, handle: str, *, title: str | None = None,
+                   why: str | None = None) -> Manifest:
+    """Announce a booth. Atomic per CLAUDE.md invariant 5: temp file +
+    os.replace, onto a temp name no other writer will pick.
+
+    OMITTED MEANS UNCHANGED; `""` MEANS CLEAR. `title` and `why` default to
+    None. The ordinary sequence is `booth new x --why "..."` then
+    `booth add x out/*.png`, and while omission meant `""` the second command
+    silently erased the sentence the first one existed to record. The shell
+    carries the distinction by leaving the environment variable UNSET rather
+    than empty.
+
+    Re-announcing PRESERVES the original `created` — `created` is when the
+    booth appeared, and saying something more about it later is not a second
+    appearance. A prior record carrying `error`, or one whose `created` is
+    `""`, is treated as having no stamp to preserve and gets `now()`: a stamp
+    that is silently wrong is worse than one that is silently new.
+
+    A WRITE THAT CHANGES NOTHING IS NOT ACTIVITY and does not touch the file,
+    so it cannot reset the booth's TTL — the rule marks learned in v0.2.0,
+    needed here because `booth link` re-announces the standing board on every
+    single post to it.
+
+    BYTES THAT COULD NOT BE READ ARE KEPT, not replaced. See INV-6.
+
+    A FAILED WRITE LEAVES NOTHING BEHIND. The temp name carries a random suffix
+    so two writers cannot share it — which also means nothing ever overwrites an
+    orphan, and `.booth.json.<hex>.tmp` is not a `.lock`, so `_newest_mtime`
+    counts it and a leak would keep a dead booth alive forever. Cleaned up on
+    every exit path.
+
+    `title` falls back to the directory name, THROUGH the same normalizer the
+    explicit value gets — a directory name may legally carry a newline on POSIX
+    and may run to 255 bytes, and the fallback used to hand either straight
+    into a card's sub-line.
+
+    Every stored string is collapsed to a single line — all runs of whitespace,
+    not only newlines, because a tab or a forty-space indent renders as badly
+    in a sub-line as a newline does — and truncated to its cap.
+
+    An empty `handle` becomes `"booth"` rather than being refused: a manifest
+    naming no handle does not read back at all, and an unreadable file is the
+    worse outcome. Unreachable from the CLI, whose fallback chain always yields
+    something; a direct caller should pass a real one.
+    """
+```
+
+## What renders
+
+One line, on both surfaces, driven by the same record. The example booth below
+is the directory `r18-ab`, announced by the handle `booth-dev`:
+
+| state | the provenance line, on an index card AND on the booth header |
+|---|---|
+| announced, with a why | `booth-dev · pick the winning denoiser` |
+| announced, no why | `booth-dev` |
+| no manifest | `unannounced` (muted) |
+| damaged manifest | `unreadable` (muted, warning tint, `title=` carries the reason) |
+
+**`title` renders too, and on exactly one surface.** An earlier draft stored it,
+surfaced a `--title` flag for it, and rendered it nowhere — a promise of a
+display name with no display, caught 4-of-4 and ranked first independently by
+every arm. It lands on the **booth page heading**, where there is room:
+`<h1>R18 A/B <span class=h1-slug>r18-ab</span></h1>`. The **index card keeps
+the directory name alone**, because that is the identity the operator navigates
+by and refers to positionally, and CLAUDE.md invariant 6 is about exactly that
+kind of reference surviving a re-render. When `title` equals the directory name
+— the default — the heading is unchanged from today.
+
+**Both index lanes get it.** The kept lane renders first and is a separate block
+in `index.html`; patching only the ephemeral lane would leave the 15 kept booths
+— the durable, most-looked-at ones — with exactly the defect this closes. This
+is the `blurtoggle` lesson (three item branches, one macro) applied to two lanes.
+
+**The booth page header gets it too**, and that is deliberate scope, not creep:
+a booth URL handed to the operator lands on the booth page, never on the index.
+Job 5 is "operator, look at this", and the page he actually opens is where the
+answer has to be.
+
+## The CLI surface
+
+Operator decision, 2026-09-22 — flags on the existing verbs, not a second verb:
+
+```sh
+booth new r18-ab --why "pick the winning denoiser"
+booth add r18-ab out/*.png --why "second pass, sharper" --title "R18 A/B"
+booth new scratch                  # still legal — handle + created, no why
+```
+
+`handle` comes from `$ALTHING_HANDLE`, falling back to `$BOOTH_SOURCE` then
+`hostname -s` — the same resolution `booth link` already uses for its rows, so
+provenance means the same thing on the board and on the card.
+
+**Nothing existing breaks.** A bare `booth new x` / `booth add x f.png` keeps
+working; the flags are optional and may sit on either side of the file
+arguments, because a glob is usually last and a flag usually after it and
+nothing enforces that. The alternative — a separate `booth announce` verb — was
+rejected because a second step is the step that gets forgotten, which is the
+69% rot's own mechanism.
+
+**A bare re-announce does not wipe what the last one said.** On a booth that has
+never announced, a bare `new`/`add` writes `{handle, created}` with no `why`. On
+one that HAS, an omitted flag leaves the stored value alone and only a supplied
+one overwrites — `--why ""` still clears, which is a different intention. This
+distinction is load-bearing rather than polite: `booth new x --why "…"` followed
+by `booth add x out/*.png` is the ordinary sequence, and the naive reading
+erases the sentence on the second command.
+
+**The handle is the CLI's three-step chain**, not `$ALTHING_HANDLE` alone:
+`${ALTHING_HANDLE:-${BOOTH_SOURCE:-$(hostname -s)}}`, identical to the one
+`booth link` already uses for its rows, so provenance means the same thing on
+the board and on the card. A session with no handle set still announces, as its
+host.
+
+## Scope — the blast-radius pass
+
+Graphify + grep, both run, because neither is sufficient alone (graphify is
+blind to function-local and DI-injected imports; grep misses transitive reach).
+
+**Every site that creates a booth directory:**
+
+| site | gets a manifest? |
+|---|---|
+| `scripts/booth new` (line 97) | yes — `$ALTHING_HANDLE` |
+| `scripts/booth add` (line 103) | yes — `$ALTHING_HANDLE` |
+| `scripts/booth link` (line 178) | yes — `handle: booth`, the standing board |
+| `app.upload` (app.py:1087) | yes — `handle: booth`, a pickup booth |
+| `marks._Locked.__enter__` (marks.py:267) | **no** — `mkdir(exist_ok=True)` on the write path; a mark written to a booth that does not exist is not an announcement, and manifest.py must not be imported by marks.py (INV-1 cross-import) |
+| `rsync` from another host | **no** — no CLI runs; this is why `unannounced` exists |
+
+**Every reader of a booth's facts:** `list_booths` (app.py:251) and `booth_view`
+— confirmed by `graphify explain list_booths` (15 edges, 4 test consumers) and
+by grep for `data_dir.iterdir` (two sites, both in app.py, both enumerating
+booths for exactly these two surfaces).
+
+**Sites that already exclude the new file and need no change**, each verified
+rather than assumed: `items.booth_items` (items.py:182), `app.zip_booth`
+(app.py:351), `marks.import_legacy_asks` (marks.py:656).
+
+**One site the first draft of this contract got WRONG, corrected by the seam
+review** (SR-1, below): the upload path's `used: set = {UPLOAD_MARKER}` filename
+dedupe set does **not** need to gain `MANIFEST_FILE`. The implementation adds it
+anyway, as consistency with the equally-unreachable entry already there, and
+says so in a comment rather than claiming it prevents anything.
+
+⚠ **Line numbers in this section are the PRE-CHANGE coordinates** the
+blast-radius pass was run against, kept because that is what makes the pass
+auditable. They have moved; `grep` the symbol, do not trust the number.
+
+## Seam review — what the real sibling surfaces said
+
+Caller-side pass against the actual modules, not against their prose. Run after
+the cold contract panel was dispatched and before any code.
+
+**SR-1 — the upload-collision change is unnecessary, and so is the one already
+there.** `safe_upload_name` (app.py) does `base = base.lstrip(".")` with the
+comment "a leading dot would hide the file from every listing", so an uploaded
+file can never be named `.booth.json` — or `.uploaded`, which means the existing
+`UPLOAD_MARKER` entry in that set has never been able to matter either. Adding
+`MANIFEST_FILE` alongside it is consistency with a redundant guard, not a fix
+for a reachable collision. Do it or don't; what the contract may not do is claim
+it prevents something. **This is the exact class the seam review exists for: a
+scope item the contract asserted from its own reasoning and the sibling's real
+surface refutes.**
+
+**SR-2 — the atomic-write pattern transfers cleanly to a dotfile, verified not
+assumed.** `marks._write_raw` derives its temp name as
+`path.with_suffix(path.suffix + ".tmp")`. For a dotfile with an extension that
+is not obviously safe — `Path(".booth.json").stem` is `".booth"`, which looks
+alarming — but `.suffix` is `".json"` and the result is `.booth.json.tmp`.
+Checked against the interpreter. The temp file is itself a dotfile, so
+`booth_items` and `zip_booth` skip it and no reader can see it mid-write.
+
+**SR-3 — the dotfile skips are on `p.name`, and all three use `rglob` or
+`iterdir` over the booth.** `items.booth_items` (items.py:182), `app.zip_booth`
+(app.py:351) and `marks.import_legacy_asks` (marks.py:656) each test
+`p.name.startswith(".")`. A manifest at the booth root is skipped by every one
+of them. Confirmed by reading the three loops, not by trusting the claim.
+
+**SR-4 — `test_stdlib_only` is parametrized `["marks", "asks", "links"]`**
+(tests/test_marks.py:279) and gains `"manifest"` as a fourth entry. The test's
+docstring calls this INV-5 while `CLAUDE.md` calls it invariant 1; that
+inconsistency predates this unit and is left alone.
+
+**SR-5 — `.booth.json` is reachable over HTTP at `/b/<name>/.booth.json`.**
+`booth_file` refuses only path escapes and non-files, not dotfiles, so a remote
+session with no filesystem access can read a booth's announcement the same way
+it already polls `/b/<n>/marks.json`. That is a feature and it is now written
+down; there is no secret in a manifest, and the Booth has no auth by design.
+
+**SR-6 — `list_booths` returns plain dicts and the templates read them by key.**
+`b.manifest` resolves through Jinja's getitem fallback. A None manifest must be
+guarded with an explicit `{% if %}` rather than relying on `b.manifest.handle`
+rendering as Undefined, because the two lanes' cards differ and a silent
+Undefined in one of them is how the kept lane would quietly keep the old defect.
+
+## Out of scope
+
+Deliberately deferred or never. Divergence here is not drift.
+
+- **A second index ordering keyed on `created`** — a "what landed" feed. Operator
+  decision, 2026-09-22: parked for v1.1. It is a new ordered collection needing
+  its own stated rule, it competes with the existing order for what "the third
+  one" means, and it has nothing to sort the 26 manifest-less booths by.
+- **`why` in the zip manifest, or a `booth ls` column.** One-liners over the
+  same record, neither on the v1 path.
+- **Enforcing that a booth MUST announce itself.** `rsync` is the documented
+  path for every host that is not nh3-dev and never runs the CLI, so a refusal
+  would break the documented workflow. The marker is the whole mechanism.
+- **Deleting, expiring or migrating anything based on the manifest.** U4 owns
+  lifetime; this unit only describes.
+- **Any change to how items, marks, blur, keep or the link board work.** The
+  manifest is a dotfile and every existing listing already skips it.
+- **Auth, or treating a manifest as trusted.** Standing non-goal; the Booth is
+  LAN-internal and a hand-written `.booth.json` is a supported input.
+- **Provenance ON a verbatim-`index.html` booth's own page.** Five live booths
+  serve the author's HTML raw, and the Booth owns no header there to put a line
+  into — it currently reaches those pages through six regexes injected into
+  arbitrary markup, which is precisely the defect U3 exists to fix. Their INDEX
+  cards carry provenance like everything else; the page itself waits for U3's
+  declared embed seam. Verified on `pewpew-ui-brief`: page renders 200, card
+  reads `unannounced`.
+
+## Invariants
+
+Numbered INV-1..5 and local to this unit. Where a repo-wide rule is meant it is
+named in words — "CLAUDE.md invariant 5", "CLAUDE.md invariant 6" — never by a
+bare number, because an earlier draft used `INV-5` for both the repo's
+atomic-write rule and this unit's render rule and the collision was caught
+3-of-4.
+
+**INV-1 — one module knows the filename.** `booth/manifest.py` is the only
+module that names `MANIFEST_FILE`. No route body, template or CLI verb opens or
+parses `.booth.json`; `write_manifest` reads it back inside that module, which
+is what INV-3 requires and is not an exception to this rule. Falsifiable and
+tested: no other file under `booth/` contains the literal `.booth.json`.
+
+**INV-2 — the read cannot raise, AND cannot cost the caller unboundedly.**
+`read_manifest` returns for every input: an absent directory, a `.booth.json`
+that is a list, a string, `null`, empty, not UTF-8, wrong-typed, missing its
+handle, nested deeply enough to overflow the parser's stack, and one larger
+than `MANIFEST_MAX_BYTES` — which is refused by `stat` before a byte is read,
+because a bound that only constrains the RETURN recreates the outage in slow
+motion. Tested per case, the size and depth cases included.
+
+**INV-3 — `created` survives re-announcement.** A second `write_manifest` on the
+same booth preserves the first `created`. A prior record carrying `error`, or
+one whose `created` is `""`, has no stamp to preserve and gets `now()`. Tested
+against a stamp that could not have come from `now()` — `_now()` is whole-second
+resolution, so back-to-back writes share a timestamp and a naive test passes
+against an implementation that regenerates it every time.
+
+**INV-4 — stdlib-only, and sibling-free** (this is CLAUDE.md invariant 1
+extended by one clause). `booth/manifest.py` imports nothing outside the
+standard library and nothing from `booth.*` — a cross-import between two
+stdlib-only modules is a second way for the repo rule to break. Relative
+imports count; the AST walk sees them.
+
+**INV-6 — bytes that could not be read are never destroyed.** When
+`write_manifest` replaces a manifest whose read returned `error`, the old bytes
+move to `QUARANTINE_FILE` first. This is the doctrine marks made explicit in
+v0.2.1 — reads lenient, writes strict, damaged bytes stay on disk — and this
+unit contradicted it by replacing outright, so a file that failed on ONE field
+lost the others with it, including a `why` the re-announcer may never have kept
+anywhere.
+
+It diverges from marks in HOW it honours the rule, and the divergence is the
+interesting part. Marks REFUSE the write and answer 409, because the operator's
+judgment is not restatable. A manifest QUARANTINES and proceeds, because
+refusing would fail `booth add` and lose the files it was mid-way through
+copying — and a booth's own description is something its poster can say again.
+One fixed quarantine name rather than a timestamped series: nothing prunes a
+booth but the sweep, and the most recent damage is the only copy anyone opens.
+
+**INV-5 — unannounced and unreadable render DIFFERENT TEXT.** Not merely
+different styling: the words differ (`unannounced` / `unreadable`), so the
+distinction survives a stylesheet change and a reader who cannot see colour. A
+one-pixel difference would satisfy a looser wording and encode nothing, and the
+point is that one of the two states is something somebody has to go and fix.

docs/contracts/u6_benches.contract.md

@@ -0,0 +1,581 @@
+---
+contract_version: "1.0"
+module: "booth.benches"
+purpose: "A bench is a running thing, registered -- not a booth, and not a bookmark. The standing link board absorbed all three jobs because only one of them had a surface, and it now carries 221 rows of which 178 (80%) are booth announcements and 156 (71% of the whole board) point at booths that were swept. U5 gave the booth announcement a home; this unit gives the RUNNING SERVICE one, and closes the loop by refusing the one shape that now has somewhere better to go. Identity is the normalized URL, so re-announcing a bench UPDATES its row instead of appending a fifth -- `talk` is on the board five times and Peedlar's root three. Nothing on the board is deleted by this unit: the dead rows are MARKED so the operator can see and remove them with the bulk control that already exists."
+depends_on:
+  - "booth.links (`booth_target` is DEFINED here and consumed there -- see INV-2. The board's existing parse/remove/pin machinery is untouched: rows keep their content-hash identity, `links.md` stays an O_APPEND multi-writer log, and no row is rewritten by anything this unit adds.)"
+  - "booth.app (the dead-row marker needs a booth-exists predicate. IT CANNOT USE `resolve_booth`: that is a CLOSURE inside `create_app`, not importable, and it RAISES HTTPException(404) -- calling it per row would turn one swept booth into a 404 for the whole board page, which is the opposite of the marker's purpose. The marker gets its own non-raising predicate carrying the SAME name-safety rules (no leading dot, no separator, no `..`) and returning False where `resolve_booth` raises. A row is dead when its target directory is absent, not when its target is nearly expired -- no new lifetime arithmetic. Verified against the real function, not assumed: seam review SR-2.)"
+language: "python"
+complexity: "medium"
+estimated_loc: 320
+confidence: 0.80
+used_by:
+  - "scripts/booth (`bench add|ls|state|rm|import` are new; `link` gains ONE refusal and is otherwise unchanged)"
+  - "booth.app.booth_view (the board's rows gain a `dead` stamp; the benches panel renders on the standing board's page)"
+  - "booth.app.list_booths (unchanged -- named here because it was checked and does NOT need to change: benches live outside the booth namespace and are invisible to it)"
+touches:
+  - "booth/benches.py (new -- the record, normalization, the lenient read, the atomic upsert, the stated order)"
+  - "booth/links.py (ONE new function, `booth_target`. No existing function changes.)"
+  - "booth/app.py (`_board_rows` stamps `dead`; the booth view passes `benches`; three POST routes for add/state/remove)"
+  - "booth/templates/booth.html (the benches panel; the dead-row marker on a board row)"
+  - "booth/templates/base.html (the .bench-* and .board-dead CSS)"
+  - "scripts/booth (the five bench verbs, the link refusal, the usage block, the header doc block)"
+  - "tests/test_benches.py (new)"
+  - "tests/test_cli.py (the bench verbs and the refusal, run against the real script under system python3)"
+  - "tests/test_marks.py (test_stdlib_only's parametrize list gains `benches`)"
+  - "docs/design/information-architecture.md (two corrections the measurement forces -- see 'What the measurement changed')"
+  - "ROADMAP.md (the bench listing order rule, which was one of the two undecided rows in the deterministic-order table)"
+assumptions:
+  - "IDENTITY IS THE FULL NORMALIZED URL, NOT THE ORIGIN, AND THIS WAS MEASURED RATHER THAN CHOSEN. Collapsing the board's 43 non-booth rows by origin yields 19 groups; by full URL, 35. The 16-group difference is not duplication -- it is EIGHT distinct gitea repositories merged into one row, THREE unrelated HuggingFace model cards merged into one, and the two LRPG surfaces on `10.100.10.50:8321` (`Authoring Studio.dc.html` and `GM Playback.dc.html`) merged into one, which are the IA doc's own example of two real benches. Origin identity would have destroyed more than it deduplicated. Full-URL identity still collapses both cases the IA doc named: `talk` 5 rows to 1, Peedlar's root 3 to 1."
+  - "THE QUERY STRING IS PART OF THE IDENTITY, the fragment is not. Measured: three ShutterChute rows differ ONLY by `?token=`, and they are three genuinely different one-shot links, not one bench posted three times -- dropping the query would merge them into a bench that is none of them. A fragment is a position inside a page, never a different resource, so it is dropped. Userinfo (`user:pass@`) is REFUSED rather than stripped: a credential must not reach a board that renders on an unauthenticated LAN surface, and silently stripping it would register a bench whose URL no longer works while telling the poster it succeeded."
+  - "`booth_target` IS HOST-AGNOSTIC AND PATH-SHAPED. A row is a booth link when its path is `/b/<name>` or `/b/<name>/...`, whatever the host. NOT a host allowlist: the fleet reaches this service as `10.100.10.50:8090`, `localhost:8090` and `nh3-dev.nh3.internal:8090`, and an allowlist would silently fail to refuse from whichever name somebody used next -- a rule that fails OPEN on the exact case it exists to catch. The accepted cost is that a third-party URL with a `/b/<x>` path would be misread; the failure is visible (a refusal naming the reason, or a row marked dead) rather than silent, and no such URL exists on the board today."
+  - "NOTHING THIS UNIT SHIPS DELETES A ROW. ROADMAP names 'a migration that deletes anything' as explicitly not in v1. `links.md` is archived verbatim before the registry is seeded, the import writes nothing without `--apply`, and the 156 dead rows are MARKED, not pruned -- removal stays the operator's two deliberate clicks through the `unlink-many` control that has existed since before this unit. The marker is what makes the existing control usable at 221 rows; it is not a second delete path."
+  - "THE SERVICE NEVER PROBES THE NETWORK. `read_benches` is a filesystem read on the render path, exactly like `read_manifest` and `marks_for`. A bench's liveness is not checked by this unit at all -- see Out of scope, where the decision and its reversal cost are stated."
+  - "`booth/benches.py` IS STDLIB-ONLY and joins the CLAUDE.md invariant 1 list, for the same reason `manifest.py` did: `scripts/booth` imports it through a `python3 -c` heredoc under the system python3 with no venv. It must also be SIBLING-FREE -- it does not import `links`, `marks`, `asks` or `manifest`, because a cross-import between two stdlib-only modules is a second way for that invariant to break. `booth_target` therefore lives in `links.py` (the board's module, where the board's callers already are) and `benches.py` does not call it; the CLI and `app.py` each import both."
+  - "THE REGISTRY IS ONE FILE AT THE DATA ROOT, `~/booth-data/.benches.json` -- a dotfile OUTSIDE the booth namespace. It is therefore not a booth, cannot be swept, cannot be mistaken for one by `list_booths` (which iterates directories), and needs no exclusion rule anywhere. Single-writer with many readers, like marks and unlike `links.md`: the operator in one browser plus CLI calls, so it is a per-file atomic replace under an flock on the read-modify-write, NOT an append log. Inheriting the append-log shape here would be the multi-writer/single-writer mistake CLAUDE.md names."
+  - "THE ON-DISK SHAPE IS AN OBJECT KEYED BY ID, not a list. Two rows with the same identity are then impossible BY CONSTRUCTION rather than by an upsert remembering to check -- which is the whole point of giving a bench an identity. The rendered order is separate and stated (INV-4); the file's key order is not load-bearing and is never read as an order."
+open_questions:
+  - "ONE BENCH, TWO URLS. `talk` is reachable as both `https://talk.nh3.phasefinal.com:8092/` (trusted cert) and `https://10.100.10.50:8092/` (internal IP, cert warning), and both are on the board with descriptions that say so. Full-URL identity correctly keeps them as two rows, because they ARE two URLs -- but they are one bench. An alias field would merge them; so would letting a bench carry a list of URLs. Neither is designed here: aliasing is a judgment about what counts as the same thing, the registry is ~14 rows, and two rows for one bench is legible. Deferred, not solved."
+  - "WHETHER `booth link` SHOULD ALSO NUDGE TOWARD `bench add` for a URL that looks like a service root. It is not refused -- measured, roughly 14 of the 35 distinct non-booth targets are reference bookmarks (repos, model cards, docs) for which the board is the right and only home, so a second refusal would break a job the board legitimately still does. A non-blocking hint is defensible and is not in this unit."
+---
+
+# U6 — benches
+
+## The defect, stated precisely
+
+Re-measured 2026-09-22 against the live board, because the numbers in the IA
+doc are a day old and the board grew:
+
+| | IA doc, 2026-09-21 | today |
+|---|---|---|
+| rows on the standing board | 211 | **221** |
+| rows that are booth URLs | not split out | **178 — 80% of the board** |
+| …whose booth no longer exists | 145 (69%) | **156 — 71% of the whole board** |
+| rows that are not booth URLs | ~40 | **43** |
+| …distinct after normalization | — | **35** |
+
+The headline number in the IA doc — *69% rot* — is **two different defects
+wearing one number**, and separating them is what makes this unit the right
+size:
+
+1. **Booth-announcement rot (178 rows).** A session posted a booth URL because
+   a booth could not announce itself. **U5 closed the cause**: a booth now
+   carries `.booth.json` and the index is the feed. Nothing yet stops the
+   habit, so the board took 11 more of these rows in the day since it was
+   measured. This unit's *enforced rule* is the stopper, and the *dead marker*
+   is what lets the operator clear what already landed.
+
+2. **Bench re-post (8 rows).** `booth link` is an append with no identity, so
+   re-announcing a bench creates a row rather than updating one: `talk` five
+   times, Peedlar's root three. This unit's *registry* is the fix, and it is
+   the smaller half — which is worth saying plainly, because the IA doc's
+   single 69% figure implies otherwise.
+
+A third thing the measurement found, which the IA doc does not describe: **the
+board has a legitimate residual job.** Of the 35 distinct non-booth targets,
+roughly 14 are running services (benches) and roughly 14 are reference
+bookmarks — gitea repositories, HuggingFace model cards, a vLLM recipe, a
+Headscale setup page. The IA doc plans for `booth link` to survive "as a
+deprecated alias". That would deprecate the only home a third of its live
+content has. **`booth link` is not deprecated by this unit.** It loses exactly
+one shape — the booth URL — and keeps the rest.
+
+## What the measurement changed
+
+Two lines of `docs/design/information-architecture.md` are wrong and are
+corrected in the same commit, rather than left for a reader to trip over:
+
+- **`id : normalized URL`** stays, but the doc does not say what normalized
+  means, and the obvious reading — the origin — is measurably destructive here
+  (8 gitea repos into one row). The doc gains the rule and the number behind it.
+- **"`booth link` … survives as a deprecated alias rather than vanishing"** is
+  struck. It survives as itself, minus one refused shape, for the reason above.
+
+## The record
+
+```python
+@dataclass(frozen=True)
+class Bench:
+    id: str        # the normalized URL — the identity, and the dict key on disk
+    url: str       # the URL AS POSTED — what a click goes to
+    name: str      # what it is
+    owner: str     # the althing handle that registered it, or "booth"
+    state: str     # "live" | "promoted" | "retired"
+    added: str     # ISO-8601 with offset, from the FIRST registration
+    updated: str   # ISO-8601 with offset, from the most recent upsert
+    error: str | None = None   # a read-time verdict; never stored
+```
+
+`id` and `url` are two fields on purpose. The identity must be normalized so
+that re-posting updates; the href must be verbatim so that a URL whose server
+cares about a trailing slash, a case-sensitive path or a query still works when
+clicked. Collapsing them would make the registry quietly change where a link
+goes, which is the kind of bug that surfaces as "the operator clicked a bench
+and got a 404" and is never traced back here.
+
+`added` survives re-registration; `updated` does not. That is the same shape as
+U5's `created`, and for the same reason: an upsert is the same bench saying
+something new about itself, not a new bench.
+
+**`updated` means the last MUTATION of the record, not the last upsert** —
+`set_bench_state` bumps it too. Amended after the cold panel read "most recent
+upsert" literally and found the code bumping on a state change: the code is
+right (a promotion is a change to the record and "last touched" should say so)
+and the earlier wording was narrower than what anyone wants the field to mean.
+
+**Caps, and what "applied" means for each — stated per field, because it is
+not the same verb for all of them.** The cold paraphrase panel found "applied at
+the write and again at the read" readable three ways (refuse / clip-for-display
+/ truncate-and-store) with a different build behind each, and 4-of-4 arms
+flagged it.
+
+| field | cap | at the write | at the read |
+|---|---|---|---|
+| `name` | 120 | **truncated** | **truncated** |
+| `owner` | 64 | **truncated** | **truncated** |
+| `url` | 2048 | **refused** (`normalize_bench_url` raises) | **damage** — reported, never clipped |
+| `state` | one of three | **refused** | **damage** |
+| `id` | 2048 | **refused**, via the url it is derived from | **not applied** — see below |
+
+`name` and `owner` are display budgets: clipping one costs a few characters in
+a panel row. **`url` is not a budget and must never be clipped**, at either end
+— INV-7 promises the click goes to the posted address byte for byte, and a
+shortened URL keeps that promise in the type system while breaking it in the
+browser. Nothing this code writes can store an over-long one; a hand-edited
+registry can, and that is damage.
+
+**`id` is capped at the WRITE ONLY, and that asymmetry is deliberate.**
+`normalize_bench_url` refuses an input over `URL_MAX`, so nothing this code
+writes can exceed it. On the read the id is the dict KEY and it is the locator
+every control posts back — `bench state`, `bench rm`, and the panel's remove
+button all address by it. Truncating a hand-edited over-long key on read would
+produce a row the operator can see and cannot act on, which is strictly worse
+than a long one. Amended after the cold panel found the code and the contract
+disagreeing here; the code was right.
+
+## Signatures
+
+```python
+BENCHES_FILE = ".benches.json"       # at the DATA ROOT — not inside a booth
+BENCH_LOCK   = ".benches.lock"
+BENCH_STATES = ("live", "promoted", "retired")
+NAME_MAX, OWNER_MAX, URL_MAX = 120, 64, 2048
+BENCHES_MAX_BYTES = 256 * 1024
+
+
+def normalize_bench_url(url: str) -> str:
+    """The identity of a bench. Raises ValueError with a reason a human can act
+    on -- the CLI prints it verbatim.
+
+    THE RULE, in full, because it is the identity and a vague identity is worse
+    than a wrong one:
+      * surrounding whitespace stripped
+      * scheme lowercased; anything but http/https is refused
+      * userinfo (`user:pass@host`) is REFUSED, never stripped
+      * host lowercased; an empty host is refused
+      * port dropped when it is the scheme default (80 for http, 443 for https)
+      * path kept verbatim, except that a bare "/" becomes ""
+      * query kept verbatim, INCLUDING its parameter order (a query is opaque)
+      * fragment dropped
+    """
+
+
+def read_benches(root: Path) -> tuple[list[Bench], str | None]:
+    """Every registered bench, in the order of `order_benches`, plus a read-time
+    error or None. NEVER RAISES -- this is on the render path (v0.2.2 lesson)."""
+
+
+def upsert_bench(root: Path, url: str, name: str, owner: str) -> tuple[Bench, bool]:
+    """Register or update by normalized URL. Returns (bench, created).
+    `added` is preserved on update; `url`, `name`, `owner`, `updated` are
+    replaced. `state` is preserved on update and is "live" on create."""
+
+
+def set_bench_state(root: Path, bench_id: str, state: str) -> Bench | None:
+    """Move a bench between live / promoted / retired. None if no such bench."""
+
+
+def remove_bench(root: Path, bench_id: str) -> Bench | None:
+    """Drop one bench. Returns the removed record, or None."""
+
+
+def order_benches(benches: Iterable[Bench]) -> list[Bench]:
+    """ORDER: (state rank, name casefolded, id) -- live before promoted before
+    retired, then alphabetical, with the id as a total tie-break so two benches
+    sharing a name cannot swap between renders. CLAUDE.md invariant 6."""
+```
+
+And in `booth/links.py`, the one addition:
+
+```python
+def booth_target(url: str) -> str | None:
+    """The booth NAME a URL points at, or None when it is not a booth URL.
+
+    ONE PREDICATE, THREE CALLERS -- the CLI's refusal, the board's dead marker,
+    and the import's classifier. They must agree: a rule that refuses a shape
+    the board then fails to mark as dead (or the reverse) is two readers of one
+    truth, which is the bug this repo has now paid for three times.
+
+    THE NAME SEGMENT IS PERCENT-DECODED. `app.py` emits booth links through
+    `quote(name, safe="")`, so a booth whose name needs encoding appears on the
+    board encoded. Comparing the raw segment against a directory name would mark
+    every such booth dead and would print the encoded form back at the poster in
+    the refusal message. Seam review SR-7.
+
+    Returns the DECODED name. A path of `/b/` with no name, or a decoded name
+    that is empty, starts with a dot, or contains a separator or `..`, is not a
+    booth link (None) — the same rules `resolve_booth` enforces, so the two
+    cannot disagree about what is addressable.
+    """
+```
+
+## The enforced rule
+
+`booth link <url>` refuses when `booth_target(url)` is not None:
+
+```
+$ booth link http://10.100.10.50:8090/b/sindra-bakeoff/ "the bakeoff"
+booth link: that is a booth, and a booth announces itself now.
+  booth new sindra-bakeoff --why "the bakeoff"     (or --why on `booth add`)
+  the index at http://10.100.10.50:8090/ is the feed.
+exit 2
+```
+
+Three properties this refusal must have, each of which is an invariant below:
+
+- **It names the alternative.** The teaching moment belongs at the point of use;
+  17 handles have the muscle memory and a bare "refused" would send them to a
+  human.
+- **It writes nothing — nothing at all.** Not the row, not the board
+  directory, not the `.booth.json` announcement `booth link` creates on first
+  use, not a lock file. The test asserts the data root's entries are unchanged,
+  not merely that `links.md` lacks the row.
+
+  *(Amended: this listed two items while INV-3 listed four, so a reader of the
+  prose alone could conclude a lock file was permissible. One list now, and it
+  is the strict one.)*
+- **It is the ONLY new refusal.** A reference bookmark is still a link.
+
+## What renders
+
+On the standing board's page, above the rows:
+
+- **The benches panel** — each bench as name, URL, owner, state, and the date
+  it was added; ordered by `order_benches`. Controls to change state and to
+  remove, both POST, both reversible in one click except remove.
+- **A board row whose booth is gone is marked dead** — visibly, with its
+  checkbox pre-reachable by the existing select-all, so the operator can tick
+  and use the `unlink-many` control already on the page. **No new delete path.**
+
+  **Dead means exactly this, and both halves are load-bearing:**
+  `booth_target(row.url)` is not None **AND** the name it returns is not a live
+  directory in the data root. A row that is not a booth link is never dead, no
+  matter what it points at — the Booth cannot know whether a gitea repo still
+  exists and must not guess. A booth link whose booth is alive is not dead. No
+  lifetime arithmetic is involved: a booth one minute from expiry is alive.
+  *(Stated after 3-of-4 cold arms read the rule two ways — predicate-driven vs
+  existence-driven — with 221 rows riding on which.)*
+
+A registry that cannot be read renders as a panel carrying its error, never as
+an absent panel and never as a 500 — the v0.2.2 lesson, which this repo learned
+by returning 500 for `/` and `/healthz` across all 25 booths.
+
+**The panel is gated on PAGE IDENTITY — the booth carries a `links.md` — and
+never on content.** A content gate (`board or benches`) hides the panel AND its
+registration form exactly when the board is empty and the registry absent,
+which is the state a fresh deployment starts in and the one where "no benches
+registered yet" is most worth saying. That is the same defect as a damaged
+panel rendering as an absent one, one level up. Amended after the cold panel
+found the content gate shipped.
+
+## The CLI surface
+
+```
+booth bench add <url> <name>      register or update; prints registered/updated
+booth bench ls                    list, in the rendered order, with ids
+booth bench state <id|url> <s>    live | promoted | retired
+booth bench rm <id|url>           remove one
+booth bench import                classify the board's rows; WRITES NOTHING
+booth bench import --apply <id>...  register ONLY the ids you name
+
+`<id|url>` takes EITHER form because the input is normalized before the lookup,
+and normalization is idempotent — an id normalizes to itself. So the id `ls`
+prints and the raw URL in the operator's scrollback both address the same row.
+Pinned by a test, because it is the property that makes the two-form promise
+true rather than merely intended.
+```
+
+`import` prints three groups — **booth rows** (skipped; `booth_target` matched),
+**candidates** (the normalized id beside the raw URL, so a collapse is visible
+before it happens), and **refused** (normalization raised, with the reason).
+
+**`--apply` REQUIRES THE IDS. A bare `--apply` is refused.** This is the
+unit's sharpest correction and it came from all four arms of the cold paraphrase
+panel independently: the first draft registered every candidate, which made the
+write path do the exact thing this document's own rationale calls impossible —
+**tell a bench from a bookmark by its URL** — silently, to roughly 14 of 35 rows
+that belong on the board. The dry run prints ids; the operator names the ones
+that are benches; an id that is not a candidate is refused and nothing is
+written. There was no selection mechanism between the report and the write, and
+the report existed precisely because the decision is not mechanizable.
+
+## The migration
+
+1. `links.md` is archived verbatim to `~/booth-data/links/links-archive-2026-09-22.md`
+   **and committed to this repo**, before anything else. Nothing the operator
+   wrote is destroyed, and the archive is version-controlled rather than living
+   only on one box.
+2. `booth bench import` proposes; the operator applies **by naming ids**.
+3. The 156 dead booth rows are marked, and removed by him or not at all.
+
+## Scope — the blast-radius pass
+
+`graphify explain` over `remove_link_entry`, `parse_link_entries`,
+`order_for_display`, `read_pins` and `toggle_pin`, cross-checked with grep
+because graphify cannot see the CLI's `python3 -c` import (it reports the
+`app.py` importers and the test callers; `scripts/booth:353` is invisible to it
+— the exact blindness CLAUDE.md names).
+
+No existing function in `links.py` changes signature or behaviour. The board's
+rows keep their content-hash identity, so every pin, every `unlink` id in the
+operator's history, and every concurrent `booth link` append keep working
+untouched.
+
+## Out of scope
+
+- **Liveness probing.** The IA doc's BENCH shape carries `last_checked` /
+  `last_ok`; ROADMAP's v1 row does not — it names *registry, identity, enforced
+  rule, migration*, and the parking lot already parks the uptime history. This
+  unit ships none of it, deliberately: it is the only part that does network
+  I/O, which is the part that reliably takes 2–5 follow-up patches for cases the
+  first shape did not anticipate — the accretion signature this whole rewrite is
+  undoing. The record is designed so adding it later is purely additive (the
+  read is lenient to unknown keys, so an older Booth reading a newer file does
+  not break). **This is a scope reduction against the IA doc and the operator
+  can reverse it; the cost of reversing it is one field pair and one CLI verb.**
+- **Pruning the board.** Not in v1, by ROADMAP.
+- **Bench aliases.** See open questions.
+- **A bench page.** A bench is a link to somewhere else; giving it a page here
+  would make the Booth a directory service.
+- **Any change to how booths announce themselves.** That was U5 and it landed.
+
+## Invariants
+
+**INV-1 — one module knows the registry's filename and shape.**
+`booth/benches.py` is the only place `.benches.json` is named, parsed or
+written. No route body and no CLI branch constructs the path or reads the JSON.
+*Falsifiable:* a test that fails if the literal `.benches.json` appears anywhere
+outside `benches.py` — and specifically fails under the change that defeats it,
+which is a route reading the file directly to save an import. Asserting only
+that the panel renders would pass under exactly that change.
+
+**INV-2 — one predicate decides what a booth URL is.** `links.booth_target` is
+the only implementation, and the CLI's refusal, the dead marker and the import's
+classifier all call it.
+*Falsifiable:* the defeating change is a second implementation — a `/b/` check
+inlined in the shell for speed, or a regex in `app.py`. One table of URLs
+(trailing slash, no slash, nested path, query, uppercase host, a non-Booth host
+with a `/b/` path, a `/b/` with no name, a percent-encoded name, a decoded `..`
+and a decoded separator) runs through the predicate, the CLI's refusal AND the
+render's dead marker.
+
+**AGREEMENT IS THE WEAKER HALF AND IS NOT THE TEST.** Three callers of one
+wrong predicate agree perfectly, so agreement alone pins nothing — the table's
+**expected values** are the independent check, and the agreement rows exist to
+catch a second implementation drifting from the first. Both are asserted; only
+one of them would survive `booth_target` itself being wrong. *(Named after a
+cold arm pointed out that the falsifier reads as though agreement were
+sufficient.)* **A bare `/b/` with no name is NOT a booth link**, and the table
+pins that.
+
+**INV-3 — a refused link writes nothing.** No row, no board directory, no
+`.booth.json`, no lock file.
+*Falsifiable:* the defeating change is moving the refusal after the `mkdir -p` /
+`announce` block in the `link` branch — which is where it would naturally land
+if written without thinking. The test refuses a link into a data root with NO
+`links` booth and asserts the directory still does not exist, not merely that
+`links.md` lacks the row. Asserting the row's absence alone would pass under the
+defeating change.
+
+**INV-4 — the rendered bench order is total and stated.** `(state rank, name
+casefolded, id)`.
+*Falsifiable:* the defeating change is dropping the `id` tie-break, which leaves
+two benches sharing a name in whatever order the dict yielded. The test
+registers two benches with the SAME name in both insertion orders and asserts
+the same output sequence from both. A test over distinct names would pass with
+no tie-break at all.
+
+**INV-5 — the read cannot raise, and cannot cost the caller unboundedly.**
+`read_benches` returns `([], "...")` for damaged, absent, oversized, or
+unreadable; it never propagates. Over `BENCHES_MAX_BYTES` is refused by size
+before it is parsed.
+*Falsifiable:* the defeating change is `json.load` without the guard. The test
+GETs the standing board's page with the registry (a) absent, (b) holding
+non-JSON bytes, (c) holding valid JSON of the wrong shape, (d) holding a
+well-formed record with a wrong-typed field, (e) over the size cap, and (f)
+chmod'd unreadable, asserting 200 for all six AND that (b)–(f) render a visible
+error rather than an empty panel. Case (d) is the one that matters: it is the
+shape that is currently 500ing the gallery elsewhere in this service.
+
+**INV-6 — the identity collapses a re-post and nothing else.** Upserting the
+same normalized URL updates one row; upserting two URLs that differ in **scheme,
+host, non-default port, path, or query** creates two. **That list is
+EXHAUSTIVE** — the only things normalization discards are a fragment, a
+scheme-default port, letter case in the scheme and host, a bare `/` path, and
+surrounding whitespace.
+
+*(Amended: this said "path, query or host" with no "only", which reads as
+illustrative and left an implementer free to "fix" the rule from the
+invariant's wording — and it omitted scheme and port, two of the five. 3-of-4
+cold arms flagged it; the falsifier now carries vectors for both.)*
+*Falsifiable:* the defeating change is normalizing to the origin. The test
+registers the eight gitea URLs measured on the live board and asserts **eight**
+benches, then registers `talk`'s five rows and asserts **one** — the same
+fixture proves both directions. A test that only checked the talk collapse would
+pass under origin normalization, which is precisely the wrong rule.
+
+**INV-7 — `url` is what a click goes to; `id` is never rendered as an href.**
+*Falsifiable:* the defeating change is rendering `bench.id` in the anchor
+because it is "the clean one". The test registers a URL whose normalization
+differs from its raw form — **an uppercase host, an explicit default port, and
+a fragment** — and asserts the anchor's `href` is the raw string, byte for byte.
+
+*(Amended: this parenthetical used to name "a trailing slash on a non-empty
+path" as one of the differences. **It is not one** — the rule list keeps a
+non-empty path verbatim, slash included, and INV-6 makes `…/p` and `…/p/` two
+benches. Two passages of this document disagreed about the same character, and
+3-of-4 cold arms found the contradiction. The rule list is correct; this
+sentence was wrong.)*
+
+**INV-8 — nothing this unit ships removes a board row.** The dead marker is a
+render-time stamp; `import` without `--apply` writes nothing anywhere; `import`
+with `--apply` writes only the registry and its lock sidecar (`.benches.json`,
+`.benches.lock`) and never touches `links.md`.
+
+*(Amended: this said "writes only `.benches.json`", which contradicted the
+unit's own assumption that every read-modify-write is held under an flock on a
+sidecar. The cold panel caught the contract arguing with itself. The
+load-bearing half — `links.md` is not touched — is unchanged and is what the
+test hashes.)*
+*Falsifiable:* the defeating change is `import --apply` "tidying up" the rows it
+consumed. The test snapshots `links.md` byte for byte, runs the full unit's CLI
+surface against it — refusal, import, import --apply, bench add, bench rm — and
+asserts the file is unchanged, including its mtime-independent content hash.
+
+**INV-9 — stdlib-only, and sibling-free.** `booth/benches.py` imports nothing
+outside the standard library and nothing from `booth.*`.
+*Falsifiable:* the defeating change is `from booth.links import booth_target` —
+which is the natural thing to write, since `booth_target` is the predicate this
+unit's CLI branch also needs.
+
+**The existing parametrized `test_stdlib_only` in tests/test_marks.py DOES
+NOT CATCH THAT, and an earlier draft of this contract claimed it did.** Its
+failure set is `{r for r in roots if r != "booth" and r not in
+sys.stdlib_module_names}` — it exempts `booth` explicitly, so a sibling import
+passes it clean. The sibling-free clause exists only in the stricter copy in
+tests/test_manifest.py. Adding `benches` to the parametrized list therefore
+buys stdlib-only and NOT sibling-free. So: `benches` joins that list AND
+`tests/test_benches.py` carries its own stricter copy, mirroring `manifest`'s,
+which fails on a `booth` root. Verified by reading the real test — seam review
+SR-1.
+
+## Seam review — what the real sibling surfaces said
+
+Run in-session against the actual `.py` files rather than their contracts,
+after the cold panel was dispatched and before any code. Seven checks, five
+findings, three of them real defects in this document. `/heid-contract-review`
+is artifact-only by design and structurally cannot run this pass: its arms read
+this file and are forbidden the siblings it borrows from.
+
+| # | seam | what the real surface said | disposition |
+|---|---|---|---|
+| **SR-1** | `test_stdlib_only` (tests/test_marks.py) | **The contract was wrong.** It claimed the parametrized test "already carries" the sibling-free clause. It does not — its failure set is `{r for r in roots if r != "booth" and ...}`, which exempts `booth` on purpose. Only tests/test_manifest.py:209 has the strict copy. | **Fixed.** INV-9 now requires both: the parametrize entry AND a stricter copy in `tests/test_benches.py`. Without this the unit would have shipped with its own INV-9 untested. |
+| **SR-2** | `resolve_booth` (booth/app.py) | **The contract invited an outage.** It named `resolve_booth` as the existence check for the dead marker. That function is a closure inside `create_app` (not importable) and **raises HTTPException(404)** — called per row, one swept booth would 404 the entire board page. It also calls `.resolve()`, a syscall per row, 178 of them on this board. | **Fixed.** `depends_on` now forbids it explicitly and specifies an own non-raising predicate with the same name-safety rules. Cost stated below. |
+| **SR-7** | `quote(name, safe="")` (app.py, booth link emission) | **The contract was silent on encoding.** Booth links are emitted percent-encoded. A `booth_target` comparing the raw path segment to a directory name marks every encoded-name booth permanently dead and echoes the encoded form back in the refusal. | **Fixed.** `booth_target` decodes, and applies `resolve_booth`'s own addressability rules so the two cannot disagree. |
+| **SR-6** | `scripts/booth` dispatch (flat `case "$cmd"`, 13 single-word verbs) | Not a defect — a gap. **`bench add` would be the first two-word verb in this script.** Nothing about the existing dispatch anticipates one, and `booth bench` with no sub-verb must not fall through into the generic usage in a way that hides which word was wrong. | **Recorded.** A nested `case` under `bench)`, and a bare `bench` prints the bench verbs specifically. Named so the implementer does not invent a third pattern. |
+| **SR-3** | `data_dir` (booth/app.py) vs `DATA` (scripts/booth) | The service resolves and expands its root in `create_app`; the CLI derives it from `$BOOTH_DATA_DIR`. Two independent derivations of one path. | **No change.** This is already true of `links.md`, `.marks.json` and `.booth.json` — pre-existing and out of this unit's scope. Recorded so it is a known property rather than a discovery. |
+| **SR-4** | `list_booths` (booth/app.py) | **Confirmed, not assumed.** `if not child.is_dir() or child.name.startswith("."): continue` — `.benches.json` fails both guards. The index cannot see the registry. | **Verified.** The assumption stands on read code. |
+| **SR-5** | `sweep_once` (booth/app.py) | **Confirmed, not assumed — and this was the dangerous one.** The sweeper iterates the data root and could in principle delete the registry. It cannot: the same `is_dir()` + leading-dot pair guards it, and `shutil.rmtree` is reached only past both. | **Verified.** Had either guard been absent this unit would have shipped a design that eats its own registry on the first tick. |
+
+**The per-render cost, stated because SR-2 surfaced it.** The dead marker runs
+once per board row: 221 rows today, 178 of which parse as booth links and cost
+one `is_dir()` each. That is one `stat` per booth row per render of the standing
+board's page — and the page already does a `booth_items` walk plus a `hold_read`
+per booth on the index, so it is not a new order of magnitude. It is bounded by
+the row count, it touches no network, and it is confined to the ONE booth that
+carries a `links.md`. If the board ever grows past a few thousand rows this
+becomes worth caching; at 221 it would be premature.
+
+## Code review — what the cold panel found
+
+`/heid-code-review` panel `01M35CK8YKEKMV7T15JXEF6A8N`, four arms, verdict
+**NOT drift-zero**. Folded in full. Three findings were independently reported
+by **all four arms**, which is the signature of a contract clause that was
+written as prose and never converted into an assertion.
+
+| # | finding | arms | disposition |
+|---|---|---|---|
+| **A** | **The panel dropped the added date.** *What renders* says "the date it was added"; `b.added` appeared nowhere in the template and no test asked for it. | 4/4 | **Fixed** — rendered, and pinned by a test. |
+| **B** | **`bench ls` printed no ids**, and the truncated URL it printed was not pasteable into `bench state\|rm`. Worse: the test's own docstring *claimed* it printed ids while asserting nothing — a claim standing in for evidence, which is how the drift would have survived CI. | 4/4 | **Fixed** — the id prints whole and last; the test now round-trips what `ls` prints back through `bench state`. |
+| **C** | **`bench import` printed the description, not the raw URL**, beside each id — hiding the five-rows-of-talk collapse the clause exists to expose. | 4/4 | **Fixed** — raw URL beside the id, description demoted to a continuation line. |
+| **D** | **An IPv6 literal lost its brackets.** `http://[::1]:8080/a` normalized to `http://::1:8080/a` — not another spelling but a BROKEN identity, so a re-post never matches the row. | 3/4 | **Fixed** — bracketed literals are re-wrapped; an *unbracketed* one is refused with a reason rather than guessed at. |
+| **H** | **INV-4's tie-break falsifier could not fail.** `_write_all` serializes with `sort_keys=True`, so both insertion orders came back off disk already id-sorted and removing the tie-break left the test green. | 1/4 | **Fixed** — the test now calls `order_benches` directly with records that tie on both prior keys. A vacuous falsifier of exactly the class `persistent-memory.d/2026-09-22-vacuous-falsifiers.md` names, found by a cold reader and not by us. |
+| **I** | **An empty board hid the whole panel**, registration form included — the state a fresh deployment starts in. | 1/4 | **Fixed** — gated on page identity. |
+| **J** | **The `booth link` refusal could fail OPEN** on a name bash's `$()` erases, because it classified by captured-text emptiness. | 1/4 | **Fixed** — the predicate answers with a `B:`/`N` sentinel, so no name can be mistaken for "not a booth". |
+| **K** | A FIFO at the registry path blocked in `open()`; a deeply-nested JSON `RecursionError` escaped the `except (ValueError, OSError)` pair. | 1/4 | **The FIFO half was already fixed** by our own pass before the reply landed. **The RecursionError half was not** — 200k open brackets is 200 KB, well inside the byte cap, and it 500'd the page the function exists to protect. Fixed. |
+| **E** | The read does not apply the `id` cap the contract promised. | 3/4 | **Contract amended, code kept.** The id is the locator every control posts back; truncating a hand-edited over-long key would make a row visible and unactionable. |
+| **F, G** | INV-5's render test covered 5 of 6 cases and asserted only status 200; INV-2's URL table never ran through the dead-marker render. | 4/4, 3/4 | **Both fixed** — the render test now covers oversized, unreadable and FIFO and asserts the error is *visible*; the full table runs through the marker. |
+
+**Also folded from the per-invariant vacuity pass** (the arms' "what would still
+pass" section, which is the single most useful thing the panel produced):
+INV-6 had no vector asserting a non-default port is part of the identity, so
+"always omit the port" passed every row; INV-3 asserted only that `links/` was
+absent, so a refusal touching any other sidecar passed; INV-8's hashed sequence
+omitted `bench ls`; INV-9's AST walk is defeated by `__import__("booth.links")`.
+All four closed.
+
+**Declined:** nothing. **Amended rather than fixed:** E, `updated`'s meaning,
+INV-8's file list, the `registered`/`created` wording, and every line number in
+this document's prose — the panel found two already stale, which is the whole
+argument against putting them in prose at all.
+
+## Bug hunt — what the cold panel found
+
+`/heid-bug-hunt` panel `01M35CRRK2RTVWWF1BN09AFQG3`, four arms, diff-scoped
+against `91fd8bc`. The most severe of the three rounds, and **three of its four
+convergent findings were already closed by our own adversarial pass before the
+reply landed** — which is the complementarity the skill claims, measured in both
+directions on one diff.
+
+| finding | arms | state when the reply landed |
+|---|---|---|
+| **A single malformed board row blanks the ENTIRE 221-row board.** `%00` in a booth name decodes to an embedded NUL; `Path.is_dir()` raises **ValueError**, not `OSError`; `_board_rows`' blanket handler returns `[]`. Every row vanishes, the page still 200s, nothing says why. | 4/4 | **Already fixed** (control-character guard). |
+| **`RecursionError` escapes `read_benches` and 500s the board page.** ~4 KB of nested brackets, well under the byte cap. **Three arms independently cited the precedent: this repo already paid for this exact class in `marks.py`** — the new module re-introduced the unguarded parse. | 4/4 | **Already fixed.** |
+| **A FIFO still blocks the render path** while the code comment claims the hang lesson was applied. | 4/4 | **Already fixed** — and the comment that lied about it was the thing that made us look. |
+| **IPv6 bracket loss.** Second independent sighting, same root. | 4/4 | **Already fixed** by the code-review round. |
+| **The benches panel is nested inside `<span class="sub">`.** A `<div>` in a `<span>`: the parser closes the span implicitly and hoists the div out, orphaning the rest of the sub-line. Nothing 500s, which is why no test could see it. | 3/4 | **OPEN — fixed now.** Moved to block level; pinned by an offset assertion and verified with a real HTML parser (0 block-in-span violations). |
+| **`_booth_exists` and `resolve_booth` disagree on a symlink.** The marker called a booth pointing outside the data root alive while the page 404s it — the row renders healthy and the link is dead. | 3/4 | **OPEN — fixed now.** Same containment, same rules. |
+| **The board append opens its fd OUTSIDE the lock.** `flock LOCK printf … >> board` reads as locked and is not: the shell opens the append fd while parsing. A concurrent `unlink` replaces the inode via `os.replace`, the old fd keeps pointing at the unlinked one, and the append **succeeds, reports success, and vanishes.** | solo | **OPEN — fixed now.** Pre-existing, not this unit's, but it is silent data loss in the file this unit lives in. Proved by holding the lock and asserting nothing is written. |
+| **A pre-planted symlink at the predictable `.benches.json.tmp.<pid>`** defeats the atomic write. The replace is atomic, not safe. | solo | **OPEN — fixed now.** `mkstemp` (O_EXCL, same directory), plus an `fsync` before the replace, because `os.replace` orders the rename and not the data behind it. |
+| A successful registration can cross the read cap and poison the registry; an empty board hides the panel. | solo | **Already fixed** by the contract round. |
+
+**Declined, with the reasoning recorded.** Kimi: the `python3 -c` guard under
+`set -e` means that on a host where `booth.links` is not importable, `booth
+link` now refuses **every** URL, not just booth ones — the refusal mechanism
+refuses everything, while the sibling `announce` call degrades gracefully.
+**True, and kept as-is deliberately.** A guard that fails open is not a guard,
+and the state it describes (the package unreachable from the script that
+computes its path from its own location) is a broken install in which `booth
+new`, `booth add` and `booth ask` are equally broken. Loud failure with a
+message naming what is missing beats silent non-enforcement. Recorded rather
+than silently dismissed, because the asymmetry with `announce` is real.
+
+**What the round says about the method.** The two lenses were complementary in
+both directions on one diff: the cold panel found three live defects the
+in-session pass missed (all three invisible to a test — a layout nesting, a
+symlink disagreement, a lock-ordering race), and the in-session pass had already
+closed three of the panel's four convergent findings. Neither substitutes for
+the other. The sharpest single line in the reply is the one noting this repo had
+already paid for the `RecursionError` class in `marks.py` — **a new module
+re-introduced a bug the codebase had a test for**, which no amount of
+reading the new module in isolation would surface.

docs/contracts/u7_navigation.contract.md

@@ -0,0 +1,226 @@
+---
+contract_version: "0.1-PROPOSED"
+status: "LANDED 2026-09-22, all four components. The operator ratified the scope departure (drop subfolder sections, add filename-prefix groups) and settled the `unanswered` open question in favour of the shipped reading. Rail, filters and grid keyboard landed at a306e2d; the groups landed in the commit carrying this revision, which also DELETED tests/test_navigation.py::test_no_group_rail_is_shipped_yet — the guard that held the departure back while the ruling was outstanding. ⚠ TWO THINGS IN THIS CONTRACT CHANGED AT IMPLEMENTATION, both measured rather than preferred: the grouping RULE (see Signatures) and INV-3, which guarded one degeneracy and needed to guard two. The original text of both is kept below, struck, because the reasoning is the useful part."
+module: "booth.items + booth.app (gallery navigation)"
+purpose: "The last unit before the 1.0 cut. A gallery booth renders as one flat wall with no way to filter it, no way to move through it from the keyboard, and no grouping — so a review of sixty-odd renders is a scroll-and-squint. ROADMAP names four components: sections, a sticky rail, filters, grid keyboard. THE MEASUREMENT KILLS THE FIRST AND REPLACES IT: not one of the eleven live gallery booths has a subdirectory, so sections buy nothing, while a filename-prefix heuristic yields 5-16 sensible groups on four of the five large galleries. This unit ships the rail, the filters, the grid keyboard, and GROUPS DERIVED FROM FILENAMES rather than from a directory tree that does not exist."
+depends_on:
+  - "booth.items.booth_items (INV-1: one resolver for item facts. `Item` gains ONE field, `group`, derived here and nowhere else. No route body derives it, exactly as no route body derives `section`, `caption` or `blurred`.)"
+  - "booth.items.Item.section (ALREADY EXISTS from U1 and STAYS. This unit does not delete it and does not render a rail from it — those are different questions. A booth that does have subdirectories keeps its section values; nothing regresses.)"
+  - "booth.app.build_gallery (the thin adapter over `booth_items`; it shapes items for the template and is where `group` reaches the page)"
+  - "booth.app.image_chain (the zoom prev/next ring. UNCHANGED, and named here because it was CHECKED: the ring is the item order filtered to images, and grouping must not reorder it -- a filter that changed what `next` means would misfile the operator's judgment, which is CLAUDE.md invariant 6's whole reason for existing.)"
+language: "python + jinja + a little javascript"
+complexity: "medium"
+estimated_loc: 300
+confidence: 0.6
+used_by:
+  - "booth.app.booth_view (the gallery page gains a rail and a filter state; the grid gains keyboard focus)"
+touches:
+  - "booth/items.py (the `group` field and its derivation)"
+  - "booth/app.py (build_gallery carries `group`; booth_view passes group counts)"
+  - "booth/templates/booth.html (the rail, the filter controls, the grid's focus affordances)"
+  - "booth/templates/base.html (rail + focus CSS)"
+  - "booth/static/embed.js (NOT TOUCHED — named because it was checked; the verbatim path has no grid)"
+  - "tests/test_items.py (group derivation)"
+  - "tests/test_navigation.py (new — rail, filters, keyboard)"
+  - "ROADMAP.md (the deterministic-order table gains the group row; U7's row is rewritten)"
+assumptions:
+  - "THE SCOPE DEPARTURE WAS RATIFIED BY THE OPERATOR 2026-09-22. ROADMAP's U7 row said `sections, rail, filters, grid keyboard`; this contract drops sections and adds filename groups. The evidence is in `persistent-memory.d/2026-09-22-u7-remeasured-before-scoping.md`: zero of eleven gallery booths have a subdirectory, the only two booths that do are reports, and `pewpew-ui-brief`'s seven subdirectories hold one image between them."
+  - "THE GROUP HEURISTIC DEGENERATES IN TWO DIRECTIONS, NOT ONE, AND THIS CONTRACT ORIGINALLY SAW ONLY THE FIRST. (a) ONE GROUP FOR EVERYTHING -- live specimen `sc-iso-spread`, `DSC0001.jpg` through `DSC0006.jpg`. (b) ONE GROUP PER ITEM -- live specimens `pewpew-ui-brief` at 23 groups for 34 items and `dfa-concepts` at 13 for 20. Both render as NO rail, because a navigation affordance that cannot navigate is worse than none: it occupies the space where the real one would be. Degeneracy (b) is the one the shipped rule actually meets on the live set, and the contract as first written would have shipped it everywhere."
+  - "GROUPING IS A VIEW, NEVER A REORDERING. The item order stays `sorted(rel)` (U1 INV-3) and the zoom ring stays that order filtered to images. Grouping and filtering change what is SHOWN and never the sequence -- so `the third one` means the same thing with a filter on as with it off, and a flag lands where the operator thinks it does. This is the whole of CLAUDE.md invariant 6 applied to a surface that did not exist when it was written."
+  - "THE PAGE WORKS WITH NO JAVASCRIPT. Filters are links with a query parameter, resolved server-side; the rail is anchors. Keyboard is the one genuinely JS-only affordance and it is additive -- the page is fully usable without it. U3 cost the verbatim path its no-JS operation and said so plainly; this unit must not quietly do the same to the gallery, which is the surface the operator actually reviews on."
+  - "VIRTUALIZATION STAYS PARKED. The largest gallery is 66 images. ROADMAP parks progressive loading with `measure the real booth before optimising it`; at this size a lazy grid is almost certainly fine, and inventing the work is the failure the parking lot exists to prevent."
+open_questions:
+  - "WHETHER THE GROUP HEURISTIC SHOULD BE OVERRIDABLE. A booth could carry a `.groups` dotfile naming its own grouping, the way `.blurred` names blur. Not designed here: no live booth wants it, the heuristic is right on four of five, and adding an override before anyone has been failed by the default is speculative. Parked, not solved."
+  - "RESOLVED 2026-09-22 — `unanswered` means `has an open pick`, the U4 hold predicate, which is what shipped. The `has no mark at all` reading is a genuinely different question and is PARKED for v1.1 rather than pending."
+---
+
+# U7 — navigation at the size the booths actually are
+
+**LANDED — all four components.**
+
+| component | ROADMAP says | state |
+|---|---|---|
+| sticky rail | ratified | **landed** — totals + per-filter counts, links not scripts |
+| filters | ratified | **landed** — all / flagged / annotated / unanswered |
+| grid keyboard | ratified | **landed** — `←/→ f n Enter Esc`, bound only when a grid exists |
+| **sections → filename groups** | **departs from it** | **landed** — ratified by the operator 2026-09-22. `test_no_group_rail_is_shipped_yet`, the guard that held it back, was deleted in the same commit that built it. |
+
+`unanswered` means **has an open pick** — the U4 hold predicate. **Settled by
+the operator 2026-09-22**; the "has no mark at all" reading is a different
+question and is parked, not pending.
+
+## The defect, re-measured rather than inherited
+
+ROADMAP sizes this unit for 270 items. **The largest gallery is now 81 items
+and 40 images.** The four booths it was written against were swept on
+2026-09-22 and the set churned again during that session. The defect is real
+and the sizing is not:
+
+| | ROADMAP's premise | measured 2026-09-22 |
+|---|---|---|
+| largest gallery | 270 images, one flat wall | **`sindra-bakeoff`, 40 images** |
+| galleries with subdirectories | "sections come from subfolders, which already exist" | **0 of 11** |
+| booths with subdirectories at all | — | 2, and **both are reports** |
+| grouping signal that does exist | — | **the filename prefix** |
+
+## Sections are dead. The prefix is not.
+
+⚠ **THE TABLE BELOW IS THE RE-MEASUREMENT, AND IT DISAGREES WITH THE ONE THIS
+CONTRACT WAS WRITTEN ON.** The original claimed the rule `strip ONE trailing
+run of digits` produced **5** groups on `sindra-bakeoff` and **1** on `sindra`.
+Neither reproduces: that rule gives **24** and **27**. The original table's own
+worked example says so out loud — it notes `00-sheet-c1-market-noon.png` has no
+trailing digit run and therefore groups as its whole stem, which makes eight of
+bakeoff's forty images eight singleton groups. **The numbers 5 and 1 are
+reproducible only by two OTHER rules** (first-two-segments gives exactly 5 on
+bakeoff; first-segment gives exactly 1 on sindra), so the table that justified
+this design was assembled from more than one heuristic. Caught by implementing
+the stated rule and running it against the live set rather than trusting the
+table beside it.
+
+**The shipped rule** — first separator-delimited segment, destemmed only when
+the stem has no separator — measured against all 17 live booths, 2026-09-22.
+`G` is groups, `med` the middle group's size, `sing` the singleton groups:
+
+| booth | items | G | med | sing | rail? |
+|---|---|---|---|---|---|
+| `sindra-corpus-v1` | 66 | 11 | 5 | 4 | **yes** — `ac 12 · bu 10 · cu 12 · fb 12 · … · wu 8` |
+| `sindra-sfw-pool` | 59 | 6 | 11 | 0 | **yes** |
+| `sindra-nude-pool` | 42 | 9 | 4 | 1 | **yes** |
+| `sindra-bakeoff` | 41 | 4 | 12 | 1 | **yes** — `00 · README · m · r`, the three real families |
+| `sindra` | 31 | 2 | 15 | 1 | **yes** |
+| `muse-clothed-repro` | 7 | 3 | 2 | 1 | **yes** — `v30`/`v35`, the axis that booth is about |
+| `pewpew-ui-brief` | 34 | 23 | 1 | 19 | no — **degeneracy (b)** |
+| `dfa-concepts` | 20 | 13 | 1 | 8 | no — **degeneracy (b)** |
+| `cr123a-to-d-sleeve` | 7 | 6 | 1 | 5 | no — degeneracy (b) |
+| `sc-iso-spread` | 6 | 1 | 6 | 0 | no — **degeneracy (a)**, `DSC0001`–`DSC0006` |
+| `music3-songs`, `krea2-lora-portability` | 3 | 1 | 3 | 0 | no — degeneracy (a) |
+| `miranda-is`, `sindra-voice-1` | 47 / 10 | 10 / 6 | 2 / 2 | 3 / 2 | **no grid at all** — both carry `index.html` and take the verbatim path |
+
+**Why the rule changed.** `strip ONE trailing run of digits` keys on the END of
+the stem, which is where the *instance number* lives — so it separates
+`m-c1-market-noon-9401` from `m-c2-rain-street-9403`, which are the same family.
+The shipped rule keys on the START, which is where the *family* lives. The
+competing heuristics measured and rejected: split-on-second-hyphen (59 groups
+from 59 files), and destemming the first segment unconditionally (merges `v30`
+with `v35`).
+
+**The honest cost.** Destemming a flat stem is what makes `ac01.png` → `ac`
+work, and it is exactly what would merge `v30` with `v35` if applied to a
+segmented name. The rule therefore has a conditional in it, which is one more
+thing than "take the first segment" — paid because `sindra-corpus-v1`, the
+largest gallery, is entirely flat names.
+
+## What ships
+
+1. **`Item.group`** — derived once, in the resolver, beside `section`.
+2. **A sticky rail** — total, per-group counts, per-filter counts, jump-to-group
+   anchors. **Absent entirely when there is one group or fewer.**
+3. **Filters** — all / flagged / annotated / unanswered, as server-resolved
+   query parameters so they work with JS off.
+4. **Grid keyboard** — `←/→` move focus, `f` flags, `n` opens a note, `Enter`
+   zooms, `Esc` clears focus. Additive; the page is complete without it.
+
+## Signatures
+
+```python
+def _group_of(rel: str) -> str | None:
+    """The grouping key for an item, or None when it has none.
+
+    THE RULE, in one line: the first separator-delimited segment of the
+    basename's stem -- with a trailing digit run stripped only when the stem has
+    no separator at all.
+
+        00-sheet-c1-market-noon.png -> 00
+        m-c1-market-noon-9401.png   -> m
+        flag-rear.png               -> flag
+        ac01.png                    -> ac     (no separator: the digits ARE it)
+        DSC0001.jpg                 -> DSC
+        v30-seed8302.png            -> v30    (separator present, so v30 != v35)
+        01.png                      -> None   (nothing before the digits)
+
+    Derived HERE and nowhere else (INV-1).
+    """
+```
+
+~~**SUPERSEDED — the rule this contract was written with.**~~ *"take the stem of
+the basename, strip ONE trailing run of digits and any single separator before
+it. `ac01.png` → `ac`; `00-sheet-c1-market-noon.png` → `00-sheet-c1-market-noon`
+(no trailing digit run, so the whole stem); `flag-rear.png` → `flag-rear`."*
+Kept struck rather than deleted: it is the rule the measurement table above was
+supposed to describe, and the mismatch between the two is the thing worth
+remembering. It keys on the end of the stem, where the instance number lives,
+and so splits families rather than gathering them.
+
+## Ordering — the rule, because invariant 6 binds
+
+| collection | rule |
+|---|---|
+| items | **unchanged** — `sorted(rel)` (U1 INV-3) |
+| the zoom ring | **unchanged** — item order filtered to images |
+| **groups among themselves** | **the position of each group's FIRST member in the RENDERED sequence** — which is `sorted(rel)` narrowed by the filter and never re-sorted. So the rail reads in the same direction the grid does, and adding a file never reshuffles the rail unless it lands first in its group. Implemented by walking `shown` once into an insertion-ordered `dict`: the walk IS the rule, so there is no second sort to drift from it. |
+| items within a group | **unchanged** — they are a filtered view of `sorted(rel)`, never re-sorted |
+| the filtered grid | **unchanged** — `sorted(rel)` with non-matching items hidden |
+
+This closes ROADMAP's outstanding U7 order question. Compare pairing is not
+this unit's problem — compare mode is parked to v1.1 with the pairing rule.
+
+## Invariants
+
+**INV-1 — one resolver derives the group.** `_group_of` is called only from
+`booth_items`. *Falsifiable:* the defeating change is a route or template
+computing a prefix inline. The test asserts no call to `_group_of` survives
+inside `create_app` — the same assertion U1 makes for `classify` and
+`render_doc`, which is why it is the shape used here.
+
+**INV-2 — grouping and filtering never reorder.** *Falsifiable:* the defeating
+change is sorting by `(group, rel)` to make the grid render contiguously, which
+looks right and silently changes what "the third one" means. The test renders a
+booth whose groups interleave in `sorted(rel)` order and asserts the rendered
+item sequence is **byte-identical** with grouping on and off, and that
+`image_chain` is unchanged under every filter.
+
+**INV-3 — a rail that cannot navigate does not render, in EITHER direction of
+degeneracy.** The rail is absent unless grouping is informative: **two or more
+groups, and the middle group holding more than one item.**
+
+- **(a) one group for everything.** Live specimen `sc-iso-spread`:
+  `DSC0001.jpg`–`DSC0006.jpg`, one group, six images. A rail with a single row
+  cannot navigate.
+- **(b) one group per item.** Live specimens `pewpew-ui-brief` (23 groups for
+  34 items) and `dfa-concepts` (13 for 20). A rail with a row per tile is a
+  second copy of the grid.
+
+*Falsifiable:* two defeating changes, each with its own test. `{% if
+rail.groups %}` in the template is true for a single group and true for N
+singletons — so the decision lives in Python, where it can be measured, and the
+template guard is the whole of it. Dropping the `>= 2` term reds
+`test_no_group_rail_when_there_is_only_one_group`; dropping the median term
+reds `test_no_group_rail_when_every_item_is_its_own_group`. Both mutations were
+RUN.
+
+~~**SUPERSEDED — INV-3 as first written.**~~ *"one group renders NO rail …the
+test uses the real `sindra`-shaped fixture (thirty files, one prefix)."* Two
+things wrong with it, and the second is why this is kept: the `sindra` fixture
+does not exist (that booth yields 27 groups under the rule stated beside it,
+and 2 under the shipped one — `sc-iso-spread` is the real specimen), and it
+guarded only degeneracy (a) when (b) is the one the live set actually
+exhibits. A contract that had shipped as written would have put a 23-row rail
+on `pewpew-ui-brief`.
+
+**INV-4 — a filter is a link, not a script.** *Falsifiable:* the defeating
+change is binding filters to a click handler. The test fetches the filtered URL
+directly and asserts the server returned the filtered grid, with no JS executed.
+
+**INV-5 — the keyboard never fires on a booth with no grid.** *Falsifiable:*
+the defeating change is binding the handler unconditionally, so `f` on the
+standing link board flags nothing and swallows the keystroke. The test asserts
+the handler is not bound when `items` is empty.
+
+## Out of scope
+
+- **Sections as a rail.** Measured worthless; `Item.section` is untouched.
+- **Compare mode.** Parked to v1.1 with its pairing rule.
+- **Virtualized loading.** Parked; measure first.
+- **A `.groups` override file.** See open questions.
+- **Anything on the verbatim path.** It has no grid.

docs/design/information-architecture.md

@@ -162,6 +162,7 @@ session that posted the set.
 ```
 BENCH
   id     : normalized URL (the identity — re-posting UPDATES, never appends)
+           NORMALIZED MEANS THE FULL URL, NOT THE ORIGIN — see below
   name   : what it is
   owner  : the agent handle that registered it
   state  : live → promoted (to Homepage) → retired
@@ -173,8 +174,51 @@ BENCH
 - `booth bench add <url> "<what>"` upserts on the normalized URL. The 5 `talk`
   rows and 4 `peedlar` rows collapse to one each, by construction.
 - **`booth link` refuses a `…:8090/b/…` URL** and names the right surface. It
-  survives as a deprecated alias rather than vanishing — 17 handles have the
+  is **not deprecated** — 17 handles have the muscle memory, the teaching moment
-  muscle memory, and the teaching moment belongs at the point of use.
+  belongs at the point of use, and (corrected 2026-09-22, U6) the board has a
+  legitimate residual job: of the 35 distinct non-booth targets on it, roughly
+  **14 are reference bookmarks** — gitea repositories, HuggingFace model cards,
+  a vLLM recipe, a Headscale setup page — for which the board is the right and
+  only home. Deprecating it would evict a third of its live content. It loses
+  exactly one shape, the booth URL, and keeps the rest.
+
+### What "normalized URL" means, and why it is not the origin
+
+Corrected 2026-09-22 while U6 was being contracted. This doc said *normalized
+URL* and left it there; the obvious reading is the origin
+(`scheme://host:port`), and that reading is **measurably destructive**.
+
+Collapsing the board's 43 non-booth rows by origin yields 19 groups; by full
+URL, 35. The 16-group difference is not duplication:
+
+| what origin identity would merge | rows |
+|---|---|
+| eight distinct gitea repositories, issues and package versions | 8 → 1 |
+| three unrelated HuggingFace model cards | 3 → 1 |
+| **the two LRPG surfaces on `10.100.10.50:8321`** — this doc's own example of two real benches | 2 → 1 |
+| two different claude.ai artifact briefs | 2 → 1 |
+
+Full-URL identity still collapses both cases this doc names — `talk` 5 rows to
+1, Peedlar's root 3 to 1 — which is the entire win, without the losses.
+
+The **query string is part of the identity** and the **fragment is not**: three
+ShutterChute rows differ only by `?token=` and are three genuinely different
+one-shot links, while a fragment is a position inside a page. Credentials in a
+URL are **refused rather than stripped** — stripping registers a bench whose URL
+no longer works while telling the poster it succeeded.
+
+### One number that was two defects
+
+This doc's headline **69% rot** is two different defects wearing one number, and
+U5 already closed the cause of the larger one:
+
+| defect | rows (2026-09-22) | what fixes it |
+|---|---|---|
+| **booth-announcement rot** — a session posts a booth URL because a booth cannot announce itself | 178 rows, 156 already dead | **U5** gave job 5 a home; U6's refusal stops the habit; U6's dead marker clears what landed |
+| **bench re-post** — an append log with no identity | 8 rows | U6's registry |
+
+Worth stating plainly because the single figure implies the registry is the big
+half. It is the smaller one.
 - Liveness is *flagged*, not enforced. A bench that stops answering gets a
   marker and a date; deleting is the operator's call. Nothing here deletes the
   operator's data on a timer.
@@ -206,21 +250,41 @@ a real DOM API. Marks land at `data-booth-mark="<id>"` anchors, which keeps the
 page*. If the line is absent, the Booth injects it at **one** insertion point, so
 every existing verbatim booth keeps working untouched.
 
-**What this deletes**, and this is the whole point of the decision:
+**What this deleted** — landed as U3, 2026-09-22:
 
-- `booth/inline.py` — 114 lines of placeholder DSL, entirely
+- `booth/inline.py` — 119 lines of placeholder DSL, entirely. One line survived:
+  `form_id`, which builds the shared `<form>` id scattered question groups bind
+  to, and which moved to `app.py` beside the route that renders them.
 - `wrap_verbatim_html` and its six regexes against arbitrary HTML
   (`_HEAD_CLOSE_RE`, `_HTML_OPEN_RE`, `_DOCTYPE_RE`, `_BODY_CLOSE_RE`,
-  `_HTML_CLOSE_RE`, `_ICON_RE`) and the doctype/charset-ordering constraints
+  `_HTML_CLOSE_RE`, `_ICON_RE`) **and both of the constraints they were
-  they are threading
+  threading.** Not satisfied more carefully — gone: nothing can displace a
-- `_BACK_CHIP`, `asks_chip` — two floating chips positioned by guessed offsets
+  leading doctype into quirks mode and nothing can push the charset `<meta>`
-- `GET /b/<name>/asks` — the standalone page that existed only because a verbatim
+  out of its detection window, because the Booth only ever APPENDS now.
-  booth could not show its own asks
+- `_BACK_CHIP`, `asks_chip` — two floating chips positioned by guessed offsets.
+  embed.js builds both in the DOM.
+- the `styles()` macro. The scoped `.bk-ask-*` rules live in embed.js next to
+  the code that mounts them, emitted once by construction instead of by a
+  seen-set.
+- `GET /b/<name>/asks` was already a 308 into `/marks` by U2; this unit left it
+  there. The standalone page it named is gone, but the URL is in the operator's
+  history and in landed reports, and a dead link teaches nothing.
 
-Regex-injecting into arbitrary author HTML is the single most fragile thing in
+**What replaced them is a substring test and a `+`.** `if EMBED_SRC not in
-the service, and it is load-bearing for the operator's most important workflow.
+html: html += EMBED_SCRIPT_TAG`. A page that declares the line is served with
+nothing added to it at all.
+
+Regex-injecting into arbitrary author HTML was the single most fragile thing in
+the service, and it was load-bearing for the operator's most important workflow.
 A declared seam costs the author one line and removes the whole class.
 
+**What it cost, stated because it is real.** The verbatim path used to work with
+no JavaScript: an ask rendered server-side and submitted through a plain form.
+It now needs the script. The guarantee that an ask is never invisible survives
+in a weaker and still-true form through surfaces that need no script — the index
+card's open-mark badge, and `/b/<name>/marks`, which renders every mark
+server-side.
+
 ---
 
 # Navigation

docs/pending/fleet-note-booth-link-refusal.md

@@ -0,0 +1,66 @@
+# PENDING — fleet note to the 17 consuming handles
+
+**Status: DRAFTED, NOT SENT.** Blocked by the auto-mode classifier on
+2026-09-22 because it is a multi-recipient send, which CLAUDE.md gates on
+explicit operator approval. The operator's blanket "accept all recs" was
+read as ratifying the note's CONTENT, not as the specific broadcast
+approval that rule requires — and the classifier agreed. Not worked around.
+
+**To send it:** the operator says go, or adds a Bash permission rule for
+`postbox send`. Recipients (17, from the live board's provenance):
+
+    hamr-dev tts-dev nh3-dev shutter-dev infra-ops comfy-dev ldp-dev
+    design-dev pewpew-dev peedlar-dev brokkr-smithy-dev draupnir
+    bifrost-dev yt-voice-clipper-dev svos-dev jackdaw-dev brokkr-scan-dev
+
+Subject: `booth: \`booth link\` now refuses a booth URL — use \`booth new --why\` instead`
+
+---
+
+ONE CHANGE THAT AFFECTS YOU, and it is a refusal you would otherwise hit
+without knowing why.
+
+`booth link` now REFUSES a booth URL.
+
+    $ booth link http://10.100.10.50:8090/b/my-run/ "the renders"
+    booth link: that is a booth, and a booth announces itself now.
+      booth new my-run --why "the renders"
+      the index at http://10.100.10.50:8090/ is the feed.
+    exit 2
+
+WHY. A booth announces itself now — `booth new` and `booth add` write a
+`.booth.json` carrying your handle and a one-line `--why`, and the index
+renders it. Posting the URL to the board on top of that creates a row that
+rots the moment the booth is swept. Measured on the live board: 178 of its 221
+rows were booth URLs and 156 of those already pointed at nothing.
+
+WHAT TO DO INSTEAD. Nothing extra — just use `--why`:
+
+    booth new my-run --why "8 renders, pick the two that hold at 4K"
+    booth add my-run out/*.png --why "..."
+
+The operator sees it on the index with your handle beside it.
+
+WHAT IS UNCHANGED. `booth link` is NOT deprecated and keeps working for
+everything else — repos, model cards, docs, recipes, any durable reference.
+Roughly 14 of the board's 35 distinct non-booth links are exactly that and the
+board is still their home. Only the booth-URL shape is refused.
+
+ALSO NEW, and optional: `booth bench add <url> <name>` registers a RUNNING
+SERVICE — your current bench, the thing that gets promoted to Homepage.
+Identity is the URL, so re-posting UPDATES the row instead of adding a fifth
+(`talk` was on the board five times). `booth bench ls` lists them.
+
+    a BOOTH   is work to review. Announces itself, swept after 24h.
+    a BENCH   is a running thing. Registered, durable, upserted by URL.
+    a LINK    is a reference bookmark. The board, unchanged.
+
+ONE MORE, since it is easy to miss: `booth link` also refuses a URL carrying
+credentials (`user:pass@host`). The board renders on an unauthenticated LAN
+surface.
+
+Shipped in booth v0.6.0/v0.6.1, deployed and live. No action needed from you
+unless you have a script that posts booth URLs to the board — that will now
+exit 2 rather than silently adding a dead row.
+
+-- booth-dev

persistent-memory.d/2026-09-21-ambiguity-pass-did-the-work.md

@@ -0,0 +1,11 @@
+# Every code-changing finding came from the AMBIGUITY pass
+
+_2026-09-21 · booth_
+
+**Every one of the panel's code-changing findings came from the
+AMBIGUITY pass, none from a paraphrase divergence** — and two arms independently
+proposed cutting the paraphrase to a drift-check for narrative-heavy contracts,
+because this contract's own frontmatter carries a plain-language narrative and the
+paraphrase was partly reading my framing back to me. That is a finding about the
+`/heid-contract-review` **skill**, not about this repo, and it was reported back
+to heid. Recorded here only so a future session does not rediscover it.

persistent-memory.d/2026-09-21-boolean-escape-hatch-as-lifetime.md

@@ -0,0 +1,11 @@
+# A boolean escape hatch as the lifetime mechanism
+
+_2026-09-21 · booth_
+
+**A boolean escape hatch as the lifetime mechanism.**
+`.forever` was added because a 24h TTL genuinely did not fit some booths —
+and then 56% of live booths ended up on it, which means it is not "ephemeral
+with an exception", it is two lifetimes wearing one lifetime's clothes, with
+the operator doing the sorting by hand. Replaced at U4 by lifetime derived
+from state (an open mark pins; viewing is activity; `keep` survives as an
+explicit reasoned pin rather than the only way to say "not yet").

persistent-memory.d/2026-09-21-deterministic-order-invariant.md

@@ -0,0 +1,16 @@
+# Deterministic order is a cross-cutting v1 invariant
+
+_2026-09-21 · booth_
+
+**Deterministic order is a cross-cutting v1 invariant** —
+operator directive, mid-implementation. Every ordered collection the Booth
+renders must have a *stated* rule producing the same sequence on every render
+of the same state; the rule can be anything defensible (byte order, time, an
+explicit number, an arbitrary-but-recorded sequence), but no rule at all is
+forbidden. It binds harder here than elsewhere because the Booth's job is
+**comparison** — the operator judges tile 47 against tile 47 and refers to
+artifacts positionally, so an order that moves between renders misfiles a flag
+or a note rather than crashing. Recorded as `ROADMAP.md` § "Cross-cutting
+invariant" (with the per-collection table) and `CLAUDE.md` invariant 6, and
+tested. Still undecided and must be settled before those units ship: **U7's
+section ordering and compare pairing**, and **U6's bench listing**.

persistent-memory.d/2026-09-21-extracted-from-eshpfi.md

@@ -0,0 +1,7 @@
+# Extracted from `eshpfi` into its own repo
+
+_2026-09-21 · booth_
+
+**Extracted from `eshpfi` into its own repo.** The accreted
+service came over whole, tests included, so `tests/test_booth.py` (1581 lines)
+is the regression net the v1 rewrite is checked against.

persistent-memory.d/2026-09-21-five-mechanisms-one-job.md

@@ -0,0 +1,13 @@
+# Five mechanisms to get one question beside one artifact
+
+_2026-09-21 · booth_
+
+**Five separate mechanisms to get one question next to one
+artifact** — `.forever`, the link board, `inline.py`'s placeholder DSL,
+`wrap_verbatim_html`'s six regexes, and the floating amber asks chip plus
+`/b/<n>/asks`. Every one is a *correct local fix* to the same global
+mismatch, which is exactly why they accumulated without anyone making a bad
+call. **The foot-gun is the sixth one:** the next "just add a small thing for
+this case" reads as reasonable and is the pattern. The git log carries the
+signature — every feature ships, then takes 2–5 patches for cases the single
+shape did not anticipate. Check the ROADMAP gate before adding a mechanism.

persistent-memory.d/2026-09-21-forever-diagnosis-is-a-prediction.md

@@ -0,0 +1,10 @@
+# The `.forever` diagnosis is a falsifiable prediction
+
+_2026-09-21 · booth_
+
+**The `.forever` diagnosis is a stated, falsifiable
+prediction.** U4 (derived lifetime) predicts the kept-rate falls to the
+genuinely-durable booths. Re-measured today: **14 of 25 booths kept (56%)**,
+against the 54% the IA doc recorded. **Re-count a fortnight after U4 lands.**
+If it does not move, the diagnosis was wrong and the boolean was doing
+something else. Tracked in the IA doc's Booth section and by this entry.

persistent-memory.d/2026-09-21-ia-and-v1-gate-landed.md

@@ -0,0 +1,11 @@
+# The information architecture and the v1 gate landed
+
+_2026-09-21 · booth_
+
+**The information architecture and the v1 gate landed**
+(`726822b`): `docs/design/information-architecture.md` names the single
+defect — *one lifetime (24h from last touch) and one shape (a folder),
+serving five jobs with different lifetimes and different shapes* — and
+`ROADMAP.md` gates v1 on seven units, each closing a **measured** defect
+rather than a wish. Both were written after a measurement pass over the live
+service, and the measurements are the load-bearing part.

persistent-memory.d/2026-09-21-jinja-hot-reload-outage.md

@@ -0,0 +1,24 @@
+# Letting Jinja hot-reload templates in the deployment root
+
+_2026-09-21 · booth_
+
+**Letting Jinja hot-reload templates while the repo is the
+deployment root** — the cause of a live outage the same day U2 landed, and the
+sharpest foot-gun in the repo. `booth.service` sets `WorkingDirectory` to this
+repo, so the running service imports these files with no build step and no
+staging copy. Python is read once at process start; Jinja's `FileSystemLoader`
+re-reads a template **on every render**. Editing `booth.html` therefore
+deployed it instantly against Python from 22:03 that knew nothing about
+`item_marks`, and **19 of 25 live booths returned 500** with
+`UndefinedError: 'item_marks' is undefined`. Neither the old code nor the new
+code was broken — the service was running both at once.
+**The lesson that generalises:** a skew between a process and the disk under it
+is invisible to the test suite by construction, so no amount of green tests
+would have caught it; the operator found it. Fixed at the source rather than
+with a reminder — the `Environment` is hand-built with `auto_reload=False`, so
+there is now ONE staleness rule (nothing takes effect until you restart) and
+the running process is always a coherent snapshot of one commit. Asserted by
+`test_templates_do_not_hot_reload_from_disk`. Watch the second-order risk the
+fix introduces: a hand-built `Environment` does not inherit `autoescape` from
+the `Jinja2Templates` constructor, and booth names, item names and mark text
+are all agent-authored strings landing in HTML.

persistent-memory.d/2026-09-21-link-board-absorbing-announce.md

@@ -0,0 +1,12 @@
+# Letting the link board absorb the announce job
+
+_2026-09-21 · booth_
+
+**Letting the link board absorb the announce job.** `booth
+link` is an `O_APPEND` write with no identity and no stated rule, so
+re-announcing a bench appends a row instead of updating one, and a booth URL
+rots the moment its booth is swept — **145 of 211 rows (69%) pointed at
+nothing**, and 22 were the same target re-posted (talk 5×, peedlar 4×). The
+rot is **structural, not drift**. The lesson that cost the most: enforcing
+the link rule without first giving the announce job a home (`.booth.json`
+provenance on the index, U5) just makes it homeless.

persistent-memory.d/2026-09-21-marks-storage-decision.md

@@ -0,0 +1,21 @@
+# Marks are one `.marks.json` per booth
+
+_2026-09-21 · booth_
+
+**Marks are stored as one `.marks.json` per booth**, atomic
+temp-file + `os.replace`, `fcntl` lock on the read-modify-write — operator
+decision, this session. Two alternatives were weighed and lost: a sidecar
+per item (`<rel>.marks.json`) and extending the existing `<stem>.ask.json`
+shape. Rationale, and the reason it is not `links.md`-shaped: **(a)** U4
+makes *"does this booth owe an answer?"* a hot question — the sweep asks it
+per booth per tick and the index asks it per card per page load, so per-item
+sidecars turn it into a full walk of all 25 booths, one of which holds 270
+files; **(b)** `links.md` is an `O_APPEND` content-hash log because **17
+agent handles write it concurrently**, whereas marks have exactly one writer
+(the operator, in one browser) and many readers — a different problem that
+must not inherit the append-log design; **(c)** `.blurred` / `.pins` /
+`.forever` already establish the per-booth dotfile as the house shape for
+operator state, and `booth_items()`'s dotfile skip means it costs nothing in
+counts, galleries or zips. Accepted cost: a corrupt `.marks.json` loses that
+booth's marks rather than one item's. Implementation deferred to U2 —
+tracked at `ROADMAP.md` U2 and by this entry.

persistent-memory.d/2026-09-21-marks-write-wiped-judgment.md

@@ -0,0 +1,15 @@
+# A write over a damaged `.marks.json` wiped the booth
+
+_2026-09-21 · booth_
+
+**A write over a damaged `.marks.json` was wiping every mark in
+the booth.** Shipped in `v0.2.0`, found by the panel (Kimi, converged with
+Hulda), fixed in `v0.2.1`. `marks_for` is deliberately lenient — unparseable
+reads as `[]` so a review page still loads — and the write path inherited that
+leniency through the same reader, so one flag click appended to an empty list and
+atomically replaced the file. The fix is an **asymmetry**, which is the reusable
+part: reads stay lenient, writes go strict (`MarksCorrupt`), damaged bytes stay
+on disk, routes answer 409 not 500. A page that renders without an annotation is
+recoverable; a file that overwrote the operator's judgment is not. Kimi also
+named the class correctly — "an author steeped in the design conversation would
+likely read past" it — and that was accurate.

persistent-memory.d/2026-09-21-partial-answer-counts-as-open.md

@@ -0,0 +1,10 @@
+# A partially-answered pick counts as OPEN
+
+_2026-09-21 · booth_
+
+**A partially-answered pick now counts as OPEN** — declared, not
+smuggled. The old index badge tested `answer is None`, so a half-answered
+four-question ask read as closed on the index while the panel beside it
+rendered `◐ partial`: the two disagreed about the same booth. Open is the
+reading that makes U4 correct — a lifetime rule that unpinned a booth on the
+first radio click would sweep a review in flight.

persistent-memory.d/2026-09-21-regex-injecting-chrome.md

@@ -0,0 +1,14 @@
+# Regex-injecting chrome into arbitrary author HTML
+
+_2026-09-21 · booth_
+
+**Regex-injecting chrome into arbitrary author HTML**
+(`wrap_verbatim_html` + `_HEAD_CLOSE_RE`, `_HTML_OPEN_RE`, `_DOCTYPE_RE`,
+`_BODY_CLOSE_RE`, `_HTML_CLOSE_RE`, `_ICON_RE`, and the doctype/charset
+ordering constraints they thread). It works today and is **still live** —
+but it is the single most fragile thing in the service and it is load-bearing
+for the operator's most important workflow. Slated for deletion at U3 in
+favour of a declared seam (`/_booth/embed.js`, mounted through a real DOM
+API), which costs an author one line and removes the whole class. Do not
+extend the regex set in the meantime; if a verbatim page breaks, that is an
+argument for U3, not for a seventh pattern.

persistent-memory.d/2026-09-21-sindra-finalists-is-the-motivation.md

@@ -0,0 +1,8 @@
+# `sindra-finalists` is U2's flag motivation, caught live
+
+_2026-09-21 · booth_
+
+**`sindra-finalists` is U2's `flag` motivation caught in the
+act** — 86 items, every one captioned, and the booth's entire name is "the
+ones the operator picked." That loop currently runs through chat, which is
+the defect `flag` closes. Evidence, not argument.

persistent-memory.d/2026-09-21-tagging-with-a-gate-in-flight.md

@@ -0,0 +1,12 @@
+# Tagging a release while a review gate was in flight
+
+_2026-09-21 · booth_
+
+**Tagging a release while a review gate was still in flight.**
+`v0.2.0` was cut and announced to 15 consuming handles; the
+`/heid-contract-review` panel — dispatched BEFORE implementation, as the
+discipline says — replied afterwards with three defects in the code that had just
+shipped, one of them silent data loss. Nothing about the tier decision was wrong;
+the *timing* was. **If a gate is outstanding on the work being released, the tag
+waits for it.** The cost was a same-hour `v0.2.1` and a correction note to peers
+who had already verified against the broken version.

persistent-memory.d/2026-09-21-tolerant-writer-over-tolerant-reader.md

@@ -0,0 +1,9 @@
+# Letting the write path share the read path's leniency
+
+_2026-09-21 · booth_
+
+**Letting the write path share the read path's leniency.** See the
+`MarksCorrupt` decision above. The general shape, worth carrying beyond marks:
+a tolerant reader and a tolerant writer over the same state are not the same
+decision, and pointing both at one function silently makes them one. Tolerate on
+read so the surface still renders; refuse on write so nothing is destroyed.

persistent-memory.d/2026-09-21-two-gates-are-complementary.md

@@ -0,0 +1,13 @@
+# Seam review and cold panel had zero overlap, twice
+
+_2026-09-21 · booth_
+
+**The two review gates are complementary, measured on one unit.**
+The caller-side **seam review** (nine findings, against the real sibling module
+surfaces) and the cold **`/heid-contract-review` panel** (four arms,
+artifact-only) had **zero overlap in both directions** on U2. The seam review
+found a scope miss the panel structurally could not see: the contract omitted
+`inline.py`, whose `place()` indexes by subscript, which a frozen dataclass
+refuses. The panel found three code defects and a missing test the seam review
+had no lens for. Matches heid's kvasir zero-overlap result on the
+conformance-versus-hunt axis. **Run both; neither substitutes.**

persistent-memory.d/2026-09-21-u2-marks-landed.md

@@ -0,0 +1,16 @@
+# U2 (marks) landed — one primitive for three mechanisms
+
+_2026-09-21 · booth_
+
+**U2 (marks) landed.** One primitive replacing three
+mechanisms. `pick` / `note` / `flag` in one `.marks.json` per booth, one read
+path (`marks_for`), one openness predicate (`open_marks`), rendered beside the
+artifact on the tile, at full size in the zoom, and in the panel. `flag` and
+`note` had no write path at all before this — the selection loop
+(`golden-candidates`, `sindra-finalists`, the `pancake-*` ladders) was running
+through chat. 242 tests. Details worth carrying: `asks.py` kept `normalize_ask`
+and gained `build_answer` (the 2026-09-09 partial-answer semantics preserved by
+moving, not rewriting) and LOST its five sidecar-storage functions;
+`GET /b/<n>/marks.json` was added because remote sessions polled
+`<stem>.answer.json` over HTTP and the sidecar's removal would have taken that
+capability with it; `/b/<n>/asks` 308s to `/marks`.

persistent-memory.d/2026-09-21-u2-seam-review-earned-it.md

@@ -0,0 +1,17 @@
+# The U2 seam review earned its place, and how
+
+_2026-09-21 · booth_
+
+**The U2 seam review earned its place, and the record should
+say how.** Nine findings against the real `booth.asks` / `booth.items` /
+`booth.inline` surfaces, two of which changed scope or behaviour: `inline.py`
+was missing from `touches` entirely (its `place()` indexes asks by
+**subscript**, which a frozen dataclass refuses — nothing else in the service
+does that), and the partial-answer inconsistency above. The cold
+`/heid-contract-review` pass is artifact-only by design and structurally
+cannot see a sibling module, so neither it nor a same-model self-review would
+have found either. Two more surfaced later and are worth the same note: a
+SECOND subscript in `inline.place` the seam review undercounted, and a
+regression in my own legacy importer that a retargeted test caught — a
+malformed sidecar that renders `⚠ broken` today would have silently vanished
+on migration.

persistent-memory.d/2026-09-21-u7-section-premise-half-wrong.md

@@ -0,0 +1,19 @@
+# U7's section premise is half wrong
+
+_2026-09-21 · booth_
+
+**U7's section premise is half wrong, and it is the half that
+matters** — found by re-measuring `~/booth-data` rather than trusting the IA
+doc. The IA says sections come from subfolders that already exist on disk;
+true, but **every booth that actually needs navigation is flat**:
+`pancake-v3-full` (270 items, 0 subfolders), `pancake-v4-full` (270, 0),
+`sindra20-engines` (98 items + 99 caption sidecars, 0), `sindra-finalists`
+(86 + 87, 0). Subfolders exist on exactly two booths — `pewpew-ui-brief` (7,
+nested to `_ds/powerpellet-design-system-<uuid>/preview`) and `dfa-concepts`
+(1) — and **both are reports**, the job where grid navigation matters least.
+So sections stay worth shipping and `Item.section` stays right, but they are
+**not** "most of the navigation fix": the rail, the filters and grid keyboard
+are all of it. Worth noting for whoever writes U7: `sindra20-engines` encodes
+its structure in the **filename prefix** (`b2-s1-<subject>-<seed>`), which is
+where a grouping heuristic would actually pay. The IA doc's claim about what
+sections buy needs a line struck — not yet edited.

persistent-memory.d/2026-09-21-v020-tagged-with-a-gate-in-flight.md

@@ -0,0 +1,14 @@
+# v0.2.0 was tagged while a gate was in flight
+
+_2026-09-21 · booth_
+
+**v0.2.0 cut and announced; v0.2.1 fixed what the announcement
+was already wrong about.** Operator approved the minor (a v1 unit closed plus a
+CLI surface change for 17 consuming handles clears the release-note bar). The
+note went to 15 handles — the 17 link-board posters minus `nh3-dev`, a host
+label, and `heid`, an oracle that does not script these verbs. Then the
+cross-frontier contract panel landed and found **three defects in the code I had
+just released**, so `v0.2.1` shipped within the hour. Sequence worth remembering:
+the release was correct by the tier bar and still premature by the discipline —
+the panel had been dispatched BEFORE implementation and its reply arrived AFTER
+the tag. **If a gate is in flight, the tag can wait for it.**

persistent-memory.d/2026-09-22-a-directive-misrouted-by-pane-title.md

@@ -0,0 +1,45 @@
+# An approved directive misrouted because pane_find addresses by a rolling title
+
+_2026-09-22 · booth_
+
+**An operator-approved directive (D-0011, sent by Miranda, telling booth-dev to
+begin U7) landed on the infra-ops handle instead.** Worth keeping for the
+mechanism, not the incident: the incident resolved cleanly and the mechanism
+did not.
+
+## What happened, and why nothing broke
+
+`pane_find` matched `terminal_2` **by its ROLLING PANE TITLE**, and that pane is
+the eshpfi-management seat rather than booth-dev. Miranda confirmed all of this
+directly when asked.
+
+infra-ops caught it and **deliberately did not relay the content as an
+instruction** — their reasoning, which is exactly right: a directive arriving as
+"infra-ops says Miranda says Vuong says" is two hops from the source, and a peer
+passing operator authority along is the thing the rules warn about. They sent a
+routing report instead, quoting only the two lines that identified the target.
+
+This session then **did not act on it**, and asked Miranda directly rather than
+taking a peer's word for the operator's. She confirmed it was genuine and
+**superseded pending the sections ruling**. Both loops closed in two messages.
+
+## The part that is still true tomorrow
+
+**A pane title that changes as work moves through the pane is not a stable
+address.** It put an approved directive on the wrong seat, and:
+
+- **the failure is silent from the sender's side.** Miranda had no signal it
+  went astray until infra-ops spoke up. A directive that misroutes to a quiet
+  or busy seat simply evaporates.
+- it landed somewhere that caught it. That was luck, not design.
+
+Reported to infra-ops as an ops matter (`01M35JJ9034E64HMA8X9C21R2N`), with the
+mechanism named and no fix proposed — not this repo's call. **Not tracked
+anywhere by booth-dev**; recorded here only so the next session does not
+re-derive it if a directive goes missing again.
+
+## The rule this confirms
+
+The CLAUDE.md Miranda exception is for Miranda relaying **directly**. A
+second-hand report of a Miranda relay is one hop too far, and infra-ops said so
+before this session had to. Going to the source cost two messages and settled it.

persistent-memory.d/2026-09-22-a-mutation-harness-that-certified-a-broken-test.md

@@ -0,0 +1,50 @@
+# A mutation harness that certified a broken test, twice, for two reasons
+
+_2026-09-22 · booth_
+
+This repo already knows that **an assertion which has never seen its own
+defeating change is not known to falsify anything** — two prior entries say so
+([[2026-09-22-vacuous-falsifiers]], [[2026-09-22-seven-of-seven-falsifiers]]).
+So U7's groups were built with a harness that applies each defeating change and
+asserts the named test goes red. **The harness itself had two defects, and both
+produce the same lie: a falsifier certified without being run.**
+
+## Defect 1 — no green baseline
+
+A test that is **already red** reports RED for every mutation thrown at it. The
+escaping test had an arithmetic slip (counted `<` against `<a`/`<nav`/`</` and
+forgot the two `<b>` elements), so it was failing for a reason unrelated to
+escaping — and the harness cheerfully reported `RED ✓ the rail markup is emitted
+with |safe`. **Run the test unmutated first; a non-zero baseline is a harness
+failure, not a proven falsifier.**
+
+## Defect 2 — the bytecode cache, which is the subtle one
+
+`if len(sizes) < 2` → `if len(sizes) < 1` is **byte-identical in size**. CPython
+validates a `.pyc` against the source's `(mtime, size)` at **one-second
+granularity** — so a mutation that lands in the same second as the revert before
+it is invisible, the cached bytecode is reused, and **the harness runs the
+unmutated code and reports the falsifier proven.**
+
+The tell was non-determinism with no cause: INV-3a certified RED on one run and
+GREEN on the next with neither the test nor the code changing, and reproduced by
+hand every time. Fix: delete `__pycache__` and set `PYTHONDONTWRITEBYTECODE=1`
+in the subprocess environment before every run.
+
+⚠ **This bites any same-size source mutation**, which is most interesting ones:
+comparison flips, off-by-one constants, `and`↔`or`, `<`↔`>`. A mutation harness
+without cache defeat is biased toward exactly the mutations most worth running.
+
+## Result
+
+12 falsifiers, 12 proved, stable across consecutive runs. Two of them only
+after these fixes — and one of the twelve (`test_group_order_is_the_position_of
+_the_first_member`) was genuinely vacuous on the first pass: its `w, x, y`
+fixture's positional order **happened to be alphabetical**, so it stayed green
+under the alphabetical-sort mutation it forbade. Rebuilt so all three plausible
+rules (position, alphabetical, count) disagree.
+
+**The harness lives in the session scratchpad and dies with the session.**
+Whether it becomes `scripts/` is an open question for the operator — this repo
+has now been bitten by vacuous falsifiers three times, and prose in a memory
+file is not an instrument.

persistent-memory.d/2026-09-22-a-wrong-shaped-answer-500s-the-gallery.md

@@ -0,0 +1,120 @@
+# A wrong-shaped answer 500s the gallery and the marks page — CLOSED 2026-09-22
+
+_2026-09-22 · booth_
+
+**Found by the U3 bug-hunt panel, measured against `42ea67f` — the commit
+BEFORE U3 — so it is not this unit's doing and was not fixed by it.** U3's own
+surface is guarded; these two are not.
+
+## The defect
+
+`.marks.json` that is **well-formed JSON with a wrong-shaped value** passes
+every reader and then raises in the renderer:
+
+```json
+{"id": "batch", "shape": "pick", "answer": {"answers": [], "notes": ""}}
+```
+
+`_hydrate` only checks `isinstance(entry.get("answer"), dict)` — it never
+validates `answer["answers"]`. So `marks_for` and `hold_read` both return the
+mark with `error = None` and **no read error at all**, and then
+`_ask_inline.html` does `a.answer.answers.get(q.key)`, Jinja asks a list for
+`.get`, and it raises `UndefinedError`.
+
+Measured, not reasoned:
+
+    PRE-U3 (42ea67f) gallery page:  500
+    PRE-U3 (42ea67f) marks page:    500
+    PRE-U3 (42ea67f) index:         200
+
+The index survives because it never renders a fragment.
+
+## Why it matters more than it looks
+
+This is **the v0.2.2 shape with a different trigger**. That outage was a
+`.marks.json` that could not be PARSED; the reader was made lenient and the
+index stopped 500ing. This one parses perfectly and breaks one layer further in,
+at render time, where no leniency exists — so the lesson "one damaged file must
+cost its own tile, not the page" is only half-implemented. `read_error` is
+answering a narrower question than every caller assumes.
+
+## What U3 did and did not do
+
+U3 added `_safe_fragments` around `_pick_fragments`, so `/b/<name>/embed.json`
+returns a per-mark `error` record instead of a 500 — the same posture
+`_hydrate_safe` takes one layer down. That protects **the verbatim path only**.
+
+`booth.html` and `marks.html` call the same macros with no such guard. Left
+alone deliberately: the gallery is named out of scope in the U3 contract, and
+widening a unit mid-flight to cover a pre-existing defect in a surface it never
+touched is the scope drift the roadmap gate exists to stop.
+
+## The design question it deserves, when it is picked up
+
+Not "wrap the other two call sites" — that is the third copy of one guard. The
+real question is **where the boundary belongs**:
+
+1. **In `_hydrate`**, validating the answer shape so a wrong-shaped answer
+   becomes `error` at hydration and every surface inherits the fix. Cleanest,
+   and consistent with declarations already being normalized on read — but it
+   widens what `error` means.
+2. **At each render site**, per-mark, as U3 did. Honest and local; three copies.
+3. **In the template**, defensively. Cheapest and worst — it hides the fact
+   that anything is wrong.
+
+(1) is the shape the rest of this module already argues for: one predicate,
+one place. Worth an operator decision because it changes what a `Mark` can be.
+
+⚠ Reproduce with the fixture in
+`tests/test_embed.py::test_a_wrongly_shaped_answer_costs_its_pick_not_the_report`,
+whose closing comment points back here.
+
+Related: [[2026-09-21-marks-write-wiped-judgment]],
+[[2026-09-22-lenient-reader-blast-radius]],
+[[2026-09-22-u3-declared-embed-seam-landed]].
+
+
+---
+
+## CLOSED — 2026-09-22, after U6, at option (1)
+
+Fixed in `_hydrate`, the option this entry argued for: **one predicate, one
+place, every surface inherits it.** The operator was asked three times where the
+guard belonged and did not answer; the placement was taken under the stated
+assumption, and it is cheap to move if he disagrees — the whole fix is one
+condition in one function.
+
+**Only the MULTI case is checked**, because only the multi case indexes: a
+single-question pick's answer IS the record, with no `answers` key to get wrong.
+Requiring one unconditionally would break every single pick — the direction a
+too-eager guard fails in, and it has its own test.
+
+Measured before and after, on the gallery booth (no `index.html`):
+
+    before   /b/g/ 500   /b/g/marks 500   / 200   /healthz 200
+    after    /b/g/ 200   /b/g/marks 200   / 200   /healthz 200
+             and the error is VISIBLE on the page, and the booth's
+             OTHER, healthy pick still renders
+
+**Two things fell out of it that are worth more than the fix.**
+
+1. **`_safe_fragments` lost its natural trigger.** Probed every wrong answer
+   shape reachable from a `.marks.json`: `answers` as a list, a string or null
+   all become hydration errors now, and a wrong-typed VALUE inside `answers`
+   renders without raising because Jinja absorbs attribute access on a
+   non-mapping. So U3's guard is now a pure backstop with **no reachable
+   natural input**. Its test was rewritten to a synthetic trigger that says so —
+   patching the shared macro module through `app.state.templates` — rather than
+   left asserting a path nothing reaches. An untested guard and a guard tested
+   by an unreachable input are the same thing.
+
+2. **The guard's own handler could not survive the failure it was handling.**
+   Building that falsifier tripped it: `_safe_fragments` caught a raising
+   `_pick_fragments` and then rebuilt the broken-ask box **through the same
+   macro module that had just raised**, so when `whole` itself was broken the
+   handler re-raised and took the whole report. Fixed, with its own test. Found
+   by accident, which is the usual way.
+
+Both new falsifiers were **verified RED against their defeating change** rather
+than assumed — the discipline from [[2026-09-22-vacuous-falsifiers]], applied to
+the fix for the entry that names it.

persistent-memory.d/2026-09-22-bug-hunt-panel.md

@@ -0,0 +1,79 @@
+# The U2 bug-hunt panel — full triage
+
+**Date:** 2026-09-22 · **Thread:** `01M33XEC1H0298C0D968FWBN7A` ·
+**Reply:** `01M33YZZ1VYGZ04JGNXNTBXDKS` · **Shipped as:** `v0.2.2`
+
+`/heid-bug-hunt` on U2's diff (+2251/−632, 20 sections, 18 post-change
+snapshots). Four arms — Gróa (Grok), Hulda (Codex), Regin (GLM-5.2), Kimi
+(kimi-k3) — artifact-only, 4/4 clean transport. Heid adjudicated **9 findings
+(6 bug / 3 robustness)**. Staleness was disclosed at build: `app.py` was edited
+after the 06:38:52Z capture.
+
+## Triage, five-category
+
+### Category 1 — genuine add (8 taken, all shipped)
+
+| # | finding | where | why it was real |
+|---|---|---|---|
+| 1 | Lock-inode split on the no-op unlink (**4/4 convergent**) | `marks._Locked` | `flock` binds to an inode; unlinking under a waiter destroys mutual exclusion silently |
+| 2 | No-op lock churn resets the TTL via **directory** mtime | `marks._Locked` + `app._newest_mtime` | the guard's own comment reasons about the lock FILE's mtime; the directory is what the sweeper reads |
+| 3 | Non-string `text` / `created` raise out of the read path | `marks._clean_text`, `marks_for` sort | `list_booths` reads every booth per page load → one bad file 500s `/` and `/healthz` |
+| 4 | Legacy import stamped `created` at whole-second resolution | `marks.import_legacy_asks` | same-second sidecars re-sorted alphabetically, reversing the order the importer had just set — violates the stated `(mtime, name)` rule |
+| 5 | `/answer` 500s on a non-string `notes` form value | `app.booth_answer` | the sibling `/note` guards it; same parser, same class of value, two answers |
+| 6 | All five mark-write routes hold a blocking `flock` on the event loop | `app.py` | a contended lock freezes every route, not just the one request |
+| 7 | CLI conflates a reader crash with "open" / "unanswered" | `scripts/booth` | `marks` printed a traceback and exited 0; `answer --wait` spun the full hour on a damaged file |
+| 8 | The inline-doc tile had `markcontrols` and not `marknotes` | `booth.html` | flag a report, cannot say why — on the one item kind that is prose |
+
+Two more taken on the same sweep, found while fixing the above rather than by
+the panel: a broken mark of any shape now renders **⚠ broken** instead of as an
+empty note (the rule `_hydrate` states for picks, applied to all three shapes),
+and the marks panel is no longer suppressed on a booth that carries a
+`links.md` *and* has marks.
+
+### Category 3 — restatement of a settled prior (1, no change)
+
+**Corrupt read → filtered writeback → silent deletion** (hulda F2, kimi F3,
+gróa F4; Heid ranked it #3). **Already fixed in `v0.2.1`** by
+`_read_raw_strict` + `MarksCorrupt` — reads lenient, writes strict. The panel
+reviewed the pre-fix capture and the staleness was disclosed up front. Verified
+against the current source before declining, not assumed.
+
+This is the exact case the cross-frontier triage discipline warns about: a
+confident, well-argued, four-arm-corroborated finding against code that no
+longer exists. **Check what the peer actually read before treating an omission
+or a defect claim as new.**
+
+### Category 4 — out of place, parked (2)
+
+- **Note-id recycling** (`note-1` reused after a withdrawal) lets a stale tab
+  delete a newer note. Real mechanism; needs two tabs and an interleaving, and
+  the Booth has one viewer. Non-reused ids are a schema change, not a patch.
+- **Unvalidated flag / note targets** accumulate orphan marks. Targets come
+  from rendered items; the operator is the only writer through the browser.
+
+### Category 5 — wrong-grounding (1)
+
+**`delete_mark` can remove a pick, not only a note.** Framed as an
+access-control divergence. There is no auth by design, and restricting it would
+remove the only way to withdraw a pick that hydrates broken. Declined; the
+docstring is the thing that was imprecise, not the behaviour.
+
+## What the round is worth remembering for
+
+1. **The two review gates stayed complementary a second time.** The contract
+   panel (2026-09-21) found three defects; this bug-hunt found eight more, with
+   **no overlap**. Both ran on the same unit. Neither substitutes.
+2. **The panel beat the code's own comments three times.** The bundle's comments
+   are unusually honest and still wrong about what protected the TTL, and
+   "written atomically" sat next to a filter-then-replace. **A comment is a
+   claim, and a claim can be tested.**
+3. **The headline bug class shipped with zero guard coverage, and both mutation
+   tables said so.** `test_a_no_op_write_does_not_touch_the_booth` asserted only
+   that `.marks.json` was absent — so removing the lock unlink, removing the
+   whole lock lifecycle, or bumping the directory clock all **SURVIVED** it. The
+   test asserted an artifact of the property instead of the property. The
+   replacement asserts `booth_age_seconds` directly, with a positive control (a
+   real mark still resets the clock) so the fix cannot overshoot into "marking
+   is never activity".
+4. **`scripts/booth` had no tests at all** and two findings lived there. It has
+   five now, running the real script under the system `python3`.

persistent-memory.d/2026-09-22-cli-exit-codes.md

@@ -0,0 +1,14 @@
+# `booth marks` / `booth answer` got real exit codes
+
+_2026-09-22 · booth_
+
+**`booth marks` / `booth answer` got real exit codes**, because
+a read that CRASHED was indistinguishable from a read that said no. `marks`
+printed a traceback and exited 0 (a caller's `jq` saw success and got
+nothing); `answer --wait` read a damaged file as "not yet" and spun for the
+full hour before blaming the operator. Now `0 ok · 1 unanswered/timed-out ·
+2 no such pick · 3 unreadable`, and `read_error()` was added to `marks.py` so
+the CLI can ask the question the browser must not: the page stays lenient, the
+machine consumer gets the truth. Also `--wait` now prints ONCE — it was
+emitting a whole JSON document per poll, so a captured `--wait` held several
+concatenated values and parsed as none of them.

persistent-memory.d/2026-09-22-doctrine-not-defect.md

@@ -0,0 +1,13 @@
+# An existing test stopped me retiring documented behaviour
+
+_2026-09-22 · booth_
+
+**An existing test stopped me retiring documented behaviour
+while fixing a race.** The mtime-restore race is real, and the clean fix —
+ignoring a booth directory's own mtime whenever the booth holds anything —
+would also have silently retired the rule that RELEASING a kept board resets
+its clock, which the CLI header, the README and a deliberately-written test
+all pin. That is a TTL doctrine change, not a bug fix. Fixed the concrete half
+(a failing `os.utime` used to escape and 500 the route), left the race stated
+in the code. **A fix that changes a documented rule is a proposal, not a
+patch.**

persistent-memory.d/2026-09-22-forever-had-a-live-positive-control.md

@@ -0,0 +1,44 @@
+# The `.forever` diagnosis got a live positive control
+
+_2026-09-22 · booth_
+
+The U4 diagnosis was that `.forever` is the only way to say three different
+things — "this is durable", "I have not answered yet", "I am still looking" —
+and that only the first is what keep means. That was an argument. **On
+2026-09-22 it stopped being one.**
+
+Census of `~/booth-data`, whole population, every value a deterministic file
+fact:
+
+| | |
+|---|---|
+| live booths | 24 |
+| carrying `.forever` | 17 (70%, up from 54% on 2026-09-21) |
+| carrying `.marks.json` at all | 4 |
+| of those, with an open pick | **4 of 4** |
+| **open pick AND `.forever`** | **3** |
+
+**Three of the four booths in the entire fleet that were waiting on an answer
+had also been pinned by hand.** That is the "not yet" case caught in the act,
+not inferred from a rate.
+
+The staleness distribution says it from the other side: **10 of the 17 kept
+booths were under one day old** — younger than the TTL, so the sentinel had
+bought them nothing and was pressed pre-emptively. Only 4 were old enough
+(2.4-4.6 days) that keep is the reason they still existed.
+
+⚠ **A number I got wrong, caught by a cross-frontier arm, kept here because the
+class repeats.** The contract first said "12 are under 1.5 days old — younger
+than the TTL". The TTL is 24 hours. 1.5 days is not younger than 24 hours. The
+measurement was sound and the sentence was not; the claim only holds at the
+one-day line, where it is 10 rather than 12. Nobody on the Claude side caught
+it, including the author twice.
+
+⚠ **The hold's live blast radius is SMALL** — only 4 booths have marks at all —
+so the `.forever` re-count prediction rests on BOTH halves of U4 and on the
+sentinel becoming unnecessary rather than forbidden. **RE-COUNT A FORTNIGHT
+AFTER U4 LANDS**, i.e. on or after **2026-10-06**. If the rate does not move,
+the honest readings are "the diagnosis was wrong" OR "the habit outlived the
+need", and a bare re-count cannot tell those apart. **The three
+open-pick-plus-`.forever` booths are the ones to watch**, because for them the
+mechanism is now unambiguous.

persistent-memory.d/2026-09-22-four-paths-to-one-fail-open-delete.md

@@ -0,0 +1,57 @@
+# Four independent paths to one fail-open delete
+
+_2026-09-22 · booth_
+
+The U4 bug-hunt panel declared invariant was **"a deletion decision must never
+be made from a read that failed"**. The panel found **four independent paths
+through it, and no single arm found all four.** That is the strongest argument
+yet for running the panel rather than one arm.
+
+1. **An entry-level hydration error lost its hold** (the round's best finding).
+   `.marks.json` parses; one mark fails normalization; `_hydrate_safe` returns a
+   `Mark` carrying `error`; `_is_open` returns False for an errored pick — **on
+   purpose**, because a broken pick can never be answered. So the booth read as
+   not-held and **swept**, while the panel beside it rendered the broken mark in
+   full. The fail-safe had been built for FILE-level damage and missed
+   ENTRY-level. A mark we cannot read is judgment we cannot see; deleting the
+   booth it belongs to is the one thing we must not do with it.
+2. **A present-but-blank `.marks.json` swept.** `_read_raw_strict` early-returns
+   for whitespace-only content — correct for the WRITE path it was written for
+   (a blank file is safe to overwrite), wrong for the DELETE path. Fixed with a
+   `blank_is_corrupt=True` flag used only by `hold_read`. ⚠ The near-regression
+   worth remembering: a **valid document with an empty `marks` list** is what
+   deleting the last mark leaves behind, and holding on THAT would make every
+   finished booth immortal. Blank bytes are damage; an empty list is an answer.
+3. **`_newest_mtime` returned 0.0 when the booth's own stat failed**, which made
+   it maximally ancient and therefore the FIRST thing the sweeper takes — a
+   permissions problem resolving to a deletion. Now returns `now`: not knowing a
+   booth's age is a reason to leave it alone. ⚠ Per-entry `FileNotFoundError`
+   stays a skip, because a dangling symlink raises it and has no mtime worth
+   counting; only OTHER stat errors mean "something is here we cannot read".
+4. **`is_kept` collapsed a stat failure into not-kept.** `Path.exists()` maps
+   ELOOP and EACCES to False. Now `lstat`, with any non-ENOENT error reading as
+   KEPT, and a `.forever` symlink counting dangling or not.
+
+**`is_held` was replaced by `hold_reason`, which returns the REASON** —
+`"open"`, `"unreadable"`, or None — rather than a bool beside a separate error
+string. Two representations of one state drift; Regin independently flagged that
+the display could not tell the two holds apart. One value, read by the sweeper
+and by all four rendering surfaces.
+
+**Convergent 3-of-4, and the one with teeth beyond lifetime:** `record_view`
+used `Path.touch()`, which FOLLOWS an existing symlink. A booth carrying a
+planted `.viewed -> /anywhere` turned every page view into an mtime write at an
+arbitrary path under the service uid — and **any fleet session can write into a
+booth, because making a folder is the whole API.** Now `os.open(..., O_NOFOLLOW)`
+plus `os.utime(fd)`; a planted link raises ELOOP into the existing swallow.
+
+⚠ **THE CAPTURE TOOLING FAILED SILENTLY AND THE PEER CAUGHT IT, NOT US.** The
+snapshot `files/` tree shipped to the arms was EMPTY. The loop was
+`for f in $IN` over a multi-line variable — and **zsh does not word-split
+unquoted parameter expansions the way bash does**, so it iterated once against a
+path that was the entire list. jekyll recovered by re-applying the bundled diff
+to HEAD and verified every file byte-identical, so the round was sound. **The
+failure mode is the dangerous one: an empty bundle reads exactly like a clean
+result.** Quote-and-split explicitly (`print -r -- $IN | while read f`) or build
+the list as a real array. Same family as `[[2026-09-22-vacuous-falsifiers]]` —
+an instrument that cannot fail loudly will fail quietly.

persistent-memory.d/2026-09-22-lenient-reader-blast-radius.md

@@ -0,0 +1,15 @@
+# The lenient reader's blast radius was the whole service
+
+_2026-09-22 · booth_
+
+**The lenient reader's blast radius was the whole service, not
+one booth.** `_clean_text` did `(text or "").replace(...)` and `marks_for`
+sorts on `(created, id)`, so a stored `text` that was a dict or a `created`
+that was a number raised out of the READ path — and `list_booths` reads every
+booth's marks on every index load. One hand-edited file 500'd `/` and
+`/healthz` for all 25 booths. Fixed in two layers, matching the house posture:
+a named type check (`_entry_type_error`) plus a `_hydrate_safe` backstop that
+cannot raise, and the panel now RENDERS an unreadable mark as ⚠ broken instead
+of as an empty note. **The general shape: a lenient reader is only lenient if
+the leniency is bounded by where it runs.** `marks_for` was written for one
+booth's page and is called in a loop over every booth.

persistent-memory.d/2026-09-22-no-notice-and-what-it-does-to-the-prediction.md

@@ -0,0 +1,54 @@
+# No fleetwide notice for U4 — and what that does to the prediction
+
+_2026-09-22 · booth_
+
+**Operator decision, 2026-09-22: do NOT tell the 17 consuming handles that
+`keep` has stopped being the way to say "waiting on an answer".** No broadcast.
+Same posture he took on U5, and the same instrument: adoption gets told apart
+from design because nobody was primed.
+
+⚠ **THIS CHANGES HOW THE 2026-10-06 RE-COUNT MUST BE READ, and a session that
+misses this will draw the wrong conclusion from a true number.**
+
+U4 has two halves and they do NOT have the same adoption cost:
+
+- **The hold rides for free.** A session that runs `booth ask` gets its booth
+  held with no knowledge of anything. The operator answering releases it. No
+  peer has to learn a thing for the mechanism to work.
+- **NOT PRESSING `keep` HAS TO BE LEARNED.** U4 makes the sentinel unnecessary
+  for the "not yet" case; it does not make it unavailable, and nothing stops a
+  habit. An agent that has always pressed `keep` while waiting will keep
+  pressing it.
+
+**So a flat `.forever` rate on 2026-10-06 does NOT falsify the diagnosis.** It
+is exactly what "the mechanism works and nobody was told" looks like — the U5
+shape, one unit later. Reporting a null result without this caveat would retire
+a correct diagnosis on the strength of an uncontrolled measurement.
+
+**Use these instead, and report all three.** The raw rate stays as context, not
+as the verdict:
+
+1. **The overlap — booths with an open pick AND `.forever`.** Was **3** on
+   2026-09-22, which is the positive control for the whole diagnosis. It falls
+   only if peers learn; it is the *adoption* number.
+
+       comm -12 <(grep -l '"shape": "pick"' ~/booth-data/*/.marks.json | xargs -n1 dirname | sort) \
+                <(dirname ~/booth-data/*/.forever | sort) | wc -l
+
+2. **Did the hold ever bind?** Count booths that were held past their TTL and
+   therefore survived a sweep they would otherwise have lost. This needs no
+   peer to change anything, so it is the honest test of whether the mechanism
+   is load-bearing at all. **If it is ZERO over a fortnight, the diagnosis was
+   wrong about VOLUME** — the "not yet" case is rarer than the sentinel rate
+   suggested — and that is a real finding. The sweeper logs what it wipes;
+   nothing yet logs what it spares, so **this counter does not exist and would
+   have to be added before it can be read.** Say so rather than guessing.
+3. **The raw `.forever` rate** — 17 of 24 (70%) on 2026-09-22. Context only,
+   now that the no-notice decision has made it a measurement of habit rather
+   than of need.
+
+⚠ **Sensitivity floor, stated because a bare "no effect" is unfalsifiable:**
+only **4 of 24** booths carried marks at all on 2026-09-22. The hold cannot
+bind on a booth with no marks, so at that population the mechanism can touch at
+most a sixth of the fleet, and an effect smaller than one or two booths is not
+resolvable by any of these counts. Related: `[[2026-09-22-forever-had-a-live-positive-control]]`.

persistent-memory.d/2026-09-22-one-number-was-two-defects.md

@@ -0,0 +1,39 @@
+# The 69% link-board rot was two defects wearing one number
+
+_2026-09-22 · booth_
+
+**Re-measuring the board before writing U6's contract split its headline number
+in half, and the half U6 owns is the smaller one.** The IA doc records *211
+rows, 145 (69%) pointing at booths that no longer exist*. Re-counted on
+2026-09-22 the board was 221 rows — and the split nobody had taken before:
+
+| | count | share |
+|---|---|---|
+| rows that are booth URLs | **178** | 80% of the board |
+| …whose booth is already swept | **156** | **71% of the whole board** |
+| rows that are NOT booth URLs | 43 | 19% |
+| …distinct after full-URL normalization | 35 | |
+| …collapsed by the re-post problem U6 names | **8 rows** | |
+
+So the 69% is:
+
+1. **Booth-announcement rot — 178 rows.** A session posted a booth URL because
+   a booth could not announce itself. **U5 already closed the cause.** Nothing
+   stopped the habit, so the board took 11 more of these in the day after it was
+   first measured.
+2. **Bench re-post — 8 rows.** An append log with no identity. This is the part
+   the registry fixes, and it is an order of magnitude smaller.
+
+**The third thing, which the IA doc does not describe at all:** of the 35
+distinct non-booth targets, roughly **14 are running services (benches)** and
+roughly **14 are reference bookmarks** — gitea repos, HuggingFace model cards, a
+vLLM recipe, a Headscale page — with the rest ephemeral one-shot links. The IA
+doc planned for `booth link` to survive "as a deprecated alias". That would have
+evicted a third of the board's live content from the only home it has. **U6 does
+not deprecate `booth link`**; it removes exactly one shape from it.
+
+**Why this is worth keeping.** The single 69% figure implies the registry is the
+big win. It is not — the enforced rule and the dead marker are. A unit scoped
+off the unsplit number would have built the registry, declared victory, and left
+178 rows rotting. Re-measure before contracting; the number in the design doc is
+a day old the moment it is written.

persistent-memory.d/2026-09-22-operator-ruled-on-the-open-five.md

@@ -0,0 +1,49 @@
+# The operator ruled on all five open items at once
+
+_2026-09-22 · booth_
+
+**"accept all recs, or make good ones, write it to handoff so I can clear."** A
+blanket ratification. Four of the five executed; one was stopped by the
+permission layer and is recorded rather than worked around.
+
+| # | item | ruling | state |
+|---|---|---|---|
+| 1 | Drop subfolder sections for filename-prefix groups | **APPROVED** | **not yet built** — the next session's first job |
+| 2 | What `unanswered` filters on | **open pick** (the shipped reading) | settled; the other reading parked to v1.1 |
+| 3 | Push `main` | **PUSH** | **DONE** — 16 commits + `v0.6.0` + `v0.6.1` now on `origin` |
+| 4 | The 17-handle althing note | send it | **BLOCKED** — see below |
+| 5 | Marks guard placement | **stays at `_hydrate`** | already there; nothing to do |
+
+## Two things the blanket ruling did NOT cover, and why
+
+**The broadcast was blocked by the auto-mode classifier, and that was right.**
+CLAUDE.md gates any multi-recipient althing send on *explicit* operator
+approval — "ask, then send, never send and report" — because the cost is
+multiplied by the recipient count and paid out of budgets the sender never
+sees. A blanket "accept all recs" ratifies the note's **content**; it is not the
+specific, informed broadcast approval that rule asks for. The classifier agreed
+and **it was not worked around**. Draft, rationale and the 17-name recipient
+list live at `docs/pending/fleet-note-booth-link-refusal.md` so they survive a
+context clear; it needs his explicit go or a `postbox send` permission rule.
+
+**"No seeding yet" survives the blanket ruling**, because it was a SPECIFIC
+prior instruction rather than a recommendation of this session's. A blanket
+acceptance of recommendations does not overwrite a direct instruction pointing
+the other way. `.benches.json` still does not exist in `~/booth-data`.
+
+## The push, recorded because it is a first
+
+`main` was **16 commits ahead** with two release tags unpushed and the whole of
+U6 single-copy on one box. Pushed with `--follow-tags`, then the two tags
+explicitly — `--follow-tags` pushed neither, because both tags are LIGHTWEIGHT
+per the SemVer policy and that flag only carries annotated ones. Worth knowing:
+**a lightweight release tag needs its own `git push origin <tag>`.**
+
+## The trap this leaves behind, and it is a real one
+
+`tests/test_navigation.py::test_no_group_rail_is_shipped_yet` was written to
+**stop an unapproved group rail from arriving by accident**. The rail is now
+approved, so that test has inverted: it will block the correct work and read
+like a genuine invariant while doing it. **Whoever builds the group rail must
+delete it in the same commit.** A guard that outlives its reason is worse than
+no guard, because the next reader trusts it.

persistent-memory.d/2026-09-22-scripts-booth-got-tests.md

@@ -0,0 +1,11 @@
+# `scripts/booth` went from zero tests to five
+
+_2026-09-22 · booth_
+
+**`scripts/booth` had zero tests and now has five**
+(`tests/test_cli.py`). The panel's guard-strength tables returned UNVERIFIED
+for every CLI claim because nothing in the suite executed the script — two of
+the round's findings lived in exactly that gap. The new tests run the real
+script under the system `python3`, which makes them a live check on INV-1
+(stdlib-only) as a side effect: a third-party import in `marks.py` now fails
+in the suite the same way it would fail on a fleet host.

persistent-memory.d/2026-09-22-seven-of-seven-falsifiers.md

@@ -0,0 +1,87 @@
+# A vacuity pass that tries the contract's own mutation agrees with itself
+
+_2026-09-22 · booth_
+
+The contract-time **vacuity pass** — for each invariant, name a change that
+defeats it and check the named test goes red — was proposed independently by
+Regin and Kimi on U4's paraphrase round, and U4's own code-review panel then
+showed **five of seven** U4 falsifiers were vacuous: a green test *cited* by an
+`INV` rather than a test that would *fail* if the invariant broke. See
+[[2026-09-22-vacuous-falsifiers]].
+
+U3 ran the pass as a real instrument rather than a promise. Script in the
+session scratchpad; for each invariant it applies the mutation the contract's
+*Falsifiable:* line names, runs the single named test, and asserts a **non-zero**
+exit, restoring the file in a `finally` either way.
+
+| INV | mutation applied | verdict |
+|---|---|---|
+| 1 declaring page untouched | append `<!-- booth -->` to the declaring branch | FALSIFIED |
+| 2 appended, never inserted | insert the tag before `<title>` instead | FALSIFIED |
+| 3 no regex on author HTML | re-declare `_ICON_RE` in `app.py` | FALSIFIED |
+| 4 openness is the server's | have `embed.js` derive open from `bk-done` | FALSIFIED |
+| 5 embed.js read once | `read_text()` per request in the route | FALSIFIED |
+| 6 tail in payload order | iterate the marks list backwards | FALSIFIED |
+| 7 unplaced questions appended | short-circuit the append branch to `if (false)` | FALSIFIED |
+
+**7/7**, and — the part that makes it a measurement rather than a ritual — an
+**unmutated control run** confirming all seven named tests are green when
+nothing is broken. Without that control, a script whose mutation silently failed
+to apply (the text not found, the wrong file) reports the same clean-looking
+table. The script halts with `MUTATION-MISS` if its target string is absent,
+for exactly that reason.
+
+## Why it is worth the ten minutes
+
+Three of the seven falsifiers are in `embed.js`, which the Python suite cannot
+see at all. INV-4, INV-6 and INV-7 are held **only** by browser tests, and
+"there is a browser test named after this invariant" is precisely the kind of
+claim that feels like coverage and can be empty. Two of those three mutations
+are one-token edits — `marks.length - 1` and `if (false)` — so the cost of
+checking was minutes and the cost of being wrong was an invariant nobody was
+holding.
+
+**The general shape:** an instrument that cannot fail loudly will fail quietly.
+Same family as the `(gasp)` tag-detection specimen in the global measurement
+rule, and as the zsh word-splitting bug that shipped an empty heid bundle —
+[[2026-09-22-four-paths-to-one-fail-open-delete]]. A clean result and a broken
+method are indistinguishable from the output alone unless something in the
+method is designed to go red.
+
+## ⚠ AND THEN THE COLD PANEL SHOWED ONE OF THE SEVEN WAS VACUOUS ANYWAY
+
+The table above is real and it was **not sufficient**. The `/heid-contract-review`
+panel (`01M351WKV666D681SSRNY7D7X6`) — three of four arms, independently —
+showed **INV-3's falsifier was vacuous**, on this contract's central promise, and
+the pass above had passed it.
+
+**Why the pass missed it.** INV-3 claims *no regular expression is applied to
+author HTML*. The test name-matched the six DELETED patterns. The mutation the
+pass applied was re-declaring `_ICON_RE` — **the pattern the contract named** —
+which the name-match caught. The mutation the invariant actually forbids is a
+regex under a *new* name (`_TAIL_RE.sub(...)` in the verbatim branch), and that
+sailed through green.
+
+> **The mutation has to come from the INVARIANT'S CLAIM, not from the
+> FALSIFIER'S EXAMPLE.** A pass that applies the contract's own suggested
+> mutation is testing the contract against itself, and it will agree.
+
+**Then the fix had a hole too, and only a re-run found it.** The repaired test
+asserts `booth/app.py` performs exactly one regex operation. Re-running the pass
+*against the fix* showed an aliased `import re as _r` routes around the call
+check under a name it does not know — still VACUOUS. Closed with an import-shape
+assertion. **Run the pass on the repair, not only on the draft.**
+
+Final state: **10/10 falsifiable**, control green, the two extra rows being the
+panel's own findings turned into falsifiers.
+
+## What this is evidence for
+
+U4 measured the problem (five of seven vacuous). U3 measured a pass working
+(7/7), then measured **the pass's own blind spot**, then measured the fix's
+blind spot. All three belong in the case if the vacuity-pass proposal is ever
+put to the operator as a `/heid*` skill amendment — and the second and third
+are the parts that stop it being adopted as a ritual that always passes.
+
+Related: [[2026-09-22-u3-declared-embed-seam-landed]],
+[[2026-09-22-the-browser-became-a-test-surface]].

persistent-memory.d/2026-09-22-size-cap-opened-a-hang.md

@@ -0,0 +1,23 @@
+# The size cap opened a service-wide hang
+
+_2026-09-22 · booth_
+
+**The U5 bug-hunt panel found a service-wide hang that the
+SIZE CAP ITSELF opened — two hours after I added the cap.** `stat` reports
+size 0 for a FIFO and 0 for a symlink to `/dev/zero`, so both sail under a
+byte cap and then `read_text` blocks with no EOF or allocates until the kernel
+intervenes. `list_booths` reads every booth on every `GET /`, so ONE such file
+stalls the front page for the whole service with no error and no recovery
+short of a restart. Reproduced (`timeout` returned 124), fixed with an
+`S_ISREG` check BEFORE the size check in both modules, verified live: the
+index answered 200 in 36 ms with two FIFOs planted. **The reusable shape:
+`st_size` answers a different question than "can this be read", and a bound
+that trusts it inherits everything it does not mean — a hardening fix opened
+a worse hole than the one it closed.** Also adopted: the upload path wrote the
+manifest ABOVE its own cleanup guard (4/4), so a failure orphaned a half-booth
+whose uniquely-named leaked temp then kept it alive forever; replace-over-
+damaged destroyed recoverable bytes (4/4, now QUARANTINED rather than refused
+— marks refuse because judgment is not restatable, a booth's description is);
+and `booth answer` spelled out its own openness test, disagreeing with
+`booth marks` about a partially-answered pick, which is a direct violation of
+U2's INV-2. Full triage in `persistent-memory.d/2026-09-22-u5-panels.md`.

persistent-memory.d/2026-09-22-the-browser-became-a-test-surface.md

@@ -0,0 +1,76 @@
+# The browser became a test surface, and the version bound is the foot-gun
+
+_2026-09-22 · booth_
+
+U3 moved load-bearing logic out of Python and into JavaScript: which fragment
+lands at which anchor, what gets appended, and whether a `<form>` scattered down
+a report still owns the controls pointing at it. **The Python suite is blind to
+every one of those.** Shipping U3 with only payload-shape tests would have
+deleted ~10 real tests and replaced them with assertions that cannot see the
+thing the operator actually depends on.
+
+So `tests/test_embed_browser.py` drives a real Chromium against a real uvicorn
+on an ephemeral port. 12 tests. It found nothing on the first run — but the
+probe that preceded it settled a design question no amount of spec-reading
+would have.
+
+## The probe, and why it had controls
+
+**Question:** if a control carrying `form="F"` is inserted into the DOM *before*
+`<form id="F">` exists, does it become that form's control? The HTML spec resets
+form owner on insertion and on the `form` attribute changing — it does NOT list
+"a matching form was inserted later". The U3 design inserts fragments in visual
+order, so this happens routinely.
+
+Four conditions, N=3 each, in Chromium 151 headless:
+
+| condition | `input.form?.id` |
+|---|---|
+| A — form inserted first (**positive control**) | `F, F, F` |
+| B — control inserted first (**the question**) | `F, F, F` |
+| C — `form="NOPE"`, no such form (**negative control**) | `null, null, null` |
+| D — remove and re-set the attribute (the proposed fix) | `F, F, F` |
+
+The positive control proves the instrument can see association at all; the
+negative proves it is not manufacturing it. Without both, B's answer means
+nothing — that is the whole lesson of
+[[2026-09-22-vacuous-falsifiers]] applied before the code instead of after.
+
+**The answer is: Chromium re-resolves it, so the fix is unnecessary there.**
+The fix shipped anyway. **Sensitivity floor: ONE ENGINE.** The operator's own
+browser was not measured, the failure mode is a form that looks filled in and
+POSTs a 400, and the guard is three lines. The measurement says "not needed
+here"; it does not say "not needed".
+
+## The foot-gun, which bit before the tests were written
+
+Browsers are **box-wide** in `/opt/ms-playwright` with
+`PLAYWRIGHT_BROWSERS_PATH` wired globally — there is no per-project
+`playwright install`. Each playwright release pins **one** Chromium revision, and
+a release wanting a revision the shared store lacks dies with:
+
+    Executable doesn't exist at /opt/ms-playwright/chromium_headless_shell-1243/…
+
+That is not a missing-dependency error and it does not name the real problem.
+The store had 1223 / 1228 / 1234; `playwright` 1.63 wanted 1243. The mapping:
+
+    1.60 -> 1223    1.61 -> 1228    1.62 -> 1234    1.63 -> 1243
+
+Hence `playwright>=1.60,<1.63` in `pyproject.toml`, **with the upper bound as the
+point** and the reason in a comment beside it. A bare `playwright` would break
+the suite on the next resolve, opaquely.
+
+## The hermeticity trade, and how it is paid
+
+A browser layer makes the suite non-hermetic — it can go red for an environment
+reason. `tests/test_embed_browser.py` therefore **skips, never fails**, when
+playwright or a usable browser is missing (`pytest.importorskip`, plus a
+`pytest.skip` on any launch failure). `pytest -q` stays green anywhere; the
+browser layer is purely additive.
+
+⚠ **The failure mode of that choice: if those 12 tests start SKIPPING on this
+box, U3's placement logic is untested and the suite still says green.** If the
+count drops from 431, check the skip reason before anything else — the pinned
+bound has probably drifted past the shared store.
+
+Related: [[2026-09-22-u3-declared-embed-seam-landed]].

persistent-memory.d/2026-09-22-third-one-branch-template-miss.md

@@ -0,0 +1,42 @@
+# The third one-branch template miss — this repo's recurring blind spot
+
+_2026-09-22 · booth_
+
+**All four arms of the U4 code-review panel found the same drift, independently.**
+That is the strongest convergence either panel has produced here.
+
+The booth header's sub-line forks on `{% if board %}`, and the U4 lifetime macro
+had been added only to the `{% else %}`. So **a booth carrying `links.md`
+rendered a link count and nothing at all about its lifetime** — no countdown, no
+hold — while INV-4 said the templates have no path that renders neither. The
+standing board being kept by construction (`booth link` drops `.forever` on
+first use) is what hid it; a **released** board or a hand-made `links.md` booth
+is a live non-kept booth on that path, and both are reachable from the UI.
+
+**This is the third of the same shape in this repo's short history:**
+
+1. `blurtoggle` — the blur only patched the image/video `<figure>`; inline docs
+   render through their OWN branch and shipped unblurred. Suite green; a live
+   look caught it.
+2. verbatim chrome — a verbatim booth's own `index.html` is served untouched, so
+   the inline marks panel never renders there. Found by looking at the live
+   service during U4, not by the suite.
+3. the board branch — this one.
+
+**The pattern: the suite renders the surface the author was thinking about.**
+Every one of these was a second branch of a conditional the author had already
+satisfied once and stopped reading. A cold reader with no idea which branch was
+"the real one" finds them; the author does not, and neither does a test the
+author wrote.
+
+**Practical consequence for this repo.** When a template gains a fact, grep the
+template for `{% if %}` in the block you edited and render EVERY branch in a
+test — one test per branch, each rendering only its own surface, or the passing
+test on branch A will mask the omission on branch B. U4 now has one per surface
+(index card, booth header, board header, marks page) for exactly this reason.
+
+Declined, and worth recording: Regin and Kimi both recommended amending INV-4 to
+carve the board header out, on the grounds that board layout belongs to U7.
+**Cutting an invariant down to fit an implementation gap is the wrong direction
+when the fix is one template edit**, and U7 owns navigation and section layout —
+not whether a header states a lifetime.

persistent-memory.d/2026-09-22-three-cold-panels-on-one-unit.md

@@ -0,0 +1,53 @@
+# Three cold panels on one unit, and what each lens could only see alone
+
+_2026-09-22 · booth_
+
+U6 ran all three `/heid*` gates plus two in-session passes. **Every one of the
+five found something the others structurally could not**, which is the
+strongest evidence this repo has for running them all rather than picking one.
+
+## The scoreboard
+
+| gate | when | found |
+|---|---|---|
+| **seam review** (in-session, sibling-aware) | before code | **3 real contract defects** — a claim about a sibling test that was false, `resolve_booth` named as a per-row predicate when it RAISES 404, and silence on percent-encoding |
+| **adversarial self-pass** (in-session) | during | **4 defects** — a FIFO hang, `unquote` leaking control characters, a fail-closed-by-accident guard, a stranded scratch file |
+| **`/heid-contract-review`** (4 arms) | parallel | **the import/apply selection gap, 4-of-4** — plus per-field cap semantics, and two passages of the document contradicting each other |
+| **`/heid-code-review`** (4 arms) | parallel | **3 surface-drift findings 4-of-4**, an IPv6 identity bug, and **a falsifier that could not fail** |
+| **`/heid-bug-hunt`** (4 arms) | parallel | a `<div>` inside a `<span>`, a symlink disagreement, an append outside its lock |
+
+## The three findings worth remembering
+
+**1. The highest-value finding was a MISSING FEATURE, and the paraphrase lens
+found it.** `bench import --apply` registered every candidate while the same
+contract said ~14 of 35 were bookmarks that must stay on the board. The dry-run
+report existed *because* the decision is not mechanizable — and then `--apply`
+ignored it. A code-vs-contract lens cannot see this: the code matched the
+contract. Only reading the contract *as prose*, for what it promises a human,
+surfaces "these two sentences cannot both be satisfied."
+
+**2. A falsifier that could not fail, again.** INV-4's tie-break test went
+through the registry, and `_write_all` serializes with `sort_keys=True` — so
+both insertion orders came back off disk already id-sorted, and removing the
+tie-break left the test green. Same class as the five vacuous U4 falsifiers.
+**We ran a vacuity pass and still shipped one**; a cold reader caught it. See
+[[2026-09-22-vacuous-falsifiers]].
+
+**3. The single sharpest line came from a cross-module memory no new-module
+review could have.** Three bug-hunt arms independently noted that **this repo
+had already paid for the `RecursionError` class in `marks.py`, with a test
+documenting it — and the new module re-introduced the unguarded parse.** No
+amount of reading `benches.py` in isolation surfaces that.
+
+## Complementarity, measured in both directions on one diff
+
+The bug-hunt panel found **three live defects the in-session pass missed** — all
+three invisible to any test (a layout nesting, a symlink disagreement, a
+lock-ordering race). The in-session pass had **already closed three of that
+panel's four convergent findings** before the reply landed. Neither substitutes
+for the other, and this round is the cleanest specimen of it so far.
+
+**One finding was declined**, with reasoning recorded in the contract: on a host
+where `booth.links` cannot be imported, `booth link` now refuses every URL
+rather than only booth ones. A guard that fails open is not a guard, and that
+state is a broken install where most of the CLI is equally broken.

persistent-memory.d/2026-09-22-two-reads-are-not-one-state.md

@@ -0,0 +1,35 @@
+# Two reads of one file are not one read of one state
+
+_2026-09-22 · booth_
+
+**The one finding across both U4 panels that changed code rather than prose,
+and it came from Hulda (Codex) on the CONTRACT-paraphrase round — before any
+code existed.**
+
+The contract specified the hold check as:
+
+    is_held(marks_for(child), read_error(child))
+
+Two reads of `.marks.json`, presented as one answer. They are not. A write or a
+repair landing between them yields a pair that described the booth at **no
+instant**, and the losing pair is `([], None)` — no marks, no error — which is
+**exactly the pair that deletes**. A lenient reader plus a strict reader, each
+correct on its own, compose into a fail-open delete.
+
+The fix is `booth.marks.hold_read(booth) -> (marks, error)`: ONE strict read
+answering both questions. `sweep_once` now does one read per booth per tick
+instead of two. And because `_read_raw_strict` **raises rather than dropping an
+entry**, a non-raising strict read returns exactly what the lenient read would —
+so the index uses that same one read for its badge too and falls back to
+`marks_for` only on the error path, where leniency is the point. Better than the
+original in both correctness and cost.
+
+**The generalisable class, in heid's words: a two-read seam presented as one
+answer is a TOCTOU race even when nothing on the page looks concurrent.** Worth
+looking for anywhere two reader functions with different strictness feed one
+decision — especially when that decision ends in `rmtree`.
+
+Related: `[[2026-09-21-marks-write-wiped-judgment]]` is the same
+reads-lenient/writes-strict asymmetry; U4 extends it to the reaper with
+"deletes strict", whose scope is **the sweeper only** — a hand delete is never
+strict, which is what gives an unreadable-marks hold an exit at all.

persistent-memory.d/2026-09-22-u2-bug-hunt-panel.md

@@ -0,0 +1,19 @@
+# The U2 bug-hunt panel was not ceremony
+
+_2026-09-22 · booth_
+
+**The U2 bug-hunt panel landed and it was not ceremony —
+`v0.2.2`.** Nine adopted findings across four arms; eight were real against
+live code and one was already fixed. The headline was **4/4 convergent from
+four different angles**: `_Locked.__exit__` unlinked `.marks.lock` on the no-op
+path, and `flock` binds to an INODE — so a writer blocked on the old inode
+proceeds while the next writer creates a fresh lock file and takes it at once.
+Two processes then run the read-modify-write concurrently and the later
+`os.replace` drops a mark, with both of them obeying the protocol. **The
+cleanup existed to protect the booth's TTL and it was failing at that too**:
+creating and removing a directory entry bumps the DIRECTORY's mtime, which is
+what `_newest_mtime` actually seeds from, so a no-op reset the clock it was
+written to leave alone. Same code region, two defects, one fix — never unlink
+the lock, exempt `.<name>.lock` dotfiles from `_newest_mtime`, and put the
+directory's mtime back after creating one. Full triage in
+`persistent-memory.d/2026-09-22-bug-hunt-panel.md`.

persistent-memory.d/2026-09-22-u3-declared-embed-seam-landed.md

@@ -0,0 +1,88 @@
+# U3 landed — the page declares the seam, the Booth mounts into it
+
+_2026-09-22 · booth_
+
+**Ten regular expressions against author-written HTML are gone.** Six in
+`wrap_verbatim_html` hunting for somewhere to hang a favicon and a chip, four in
+`booth/inline.py` substituting rendered ask markup into the author's own tags.
+What replaced them, in full:
+
+```python
+return html if declares_embed(html) else html + EMBED_SCRIPT_TAG
+```
+
+A substring test and a `+`. **Both of the old wrapper's hard constraints stopped
+existing rather than being satisfied more carefully** — nothing can displace a
+leading doctype into quirks mode and nothing can push the charset `<meta>` out
+of its first-1024-byte window, because nothing in front of them ever moves.
+
+## What moved where
+
+| was | is |
+|---|---|
+| `wrap_verbatim_html` + 6 regexes | `embed_verbatim` — one `in`, one `+` |
+| `booth/inline.py`, 119 lines | deleted; `form_id` survived into `app.py` |
+| `_BACK_CHIP`, `asks_chip` | built in the DOM by `embed.js` |
+| `inject_asks` | `GET /b/<name>/embed.json` + placement in `embed.js` |
+| `_ask_inline.html`'s `styles()` | the CSS lives in `embed.js` |
+| `FAVICON_LINK` string injection | `document.querySelector('link[rel~="icon"]')` |
+
+**The fragments are still rendered by Jinja.** `embed.js` places what comes back
+and never builds one — a second renderer in JavaScript would be the same bug
+INV-1 exists to stop, in a new language. The payload also decides openness
+(`open_marks`) and order, so the page has no opinion about either.
+
+## The thing the contract got wrong, and the seam review caught
+
+The payload first keyed `questions` by question key. **A single-question pick
+normalizes to `questions: [{"key": None, …}]`** (`asks.normalize_ask`, the
+`multi: False` branch), and `json.dumps` writes that key as the string `"null"`
+— inventing a name that collides with a real key. Every one-question ask in the
+fleet would have hit it, including the live `sindra-voice-1`. `questions` is a
+LIST of `{key, html}` now; the key is nullable, and declaration order rides in
+the format instead of leaning on object-key insertion order.
+
+The cold contract panel could not have found this: it is a fact about
+`booth/asks.py`, which an artifact-only reader never sees. Third time the seam
+review has caught what the cold pass structurally cannot — see
+[[2026-09-21-two-gates-are-complementary]].
+
+## The live report that was already subtly broken
+
+`dfa-concepts/index.html` writes `<div class="ask" data-booth-ask="dfa:logo">
+<h3>The one asset that must survive</h3>`. `_EL_RE` matched the **opening tag**
+and replaced it, so the author's `.ask` wrapper class vanished, the heading was
+orphaned and the `</div>` went stray. Nobody filed a bug, because a page that is
+95% right does not look broken.
+
+`el.insertAdjacentHTML("beforeend", frag)` keeps the element and its contents
+and puts the fragment inside. Verified live in a real browser: 5 author `.ask`
+wrappers intact, 5 headings intact, 14 radios mounted inside them, zero console
+errors. **The replacement is not just less fragile, it renders the operator's
+own report more faithfully than the thing it replaced.**
+
+## The cost, stated because it is real
+
+The verbatim path used to work with **no JavaScript** — server-rendered ask, plain
+form POST, HTML5 `form=` binding resolved at parse time. It needs the script now.
+The operator's 2026-09-21 ruling accepts that; this entry records the consequence
+so nobody meets it as a surprise. The never-invisible guarantee survives in a
+weaker and still-true form through surfaces needing no script: the index card's
+open-mark badge, and `/b/<name>/marks`.
+
+## Anchor syntax
+
+`data-booth-mark` is canonical (U2 made an ask one shape of mark).
+`data-booth-ask` is a kept alias — 2 of the 4 live verbatim booths spell it that
+way, in the operator's own reports, and the alias is one clause in one selector
+string. The `<!-- booth:ask … -->` comment forms were **dropped, not ported**:
+zero users across all 21 live booths, and a page that used one falls back to the
+append path, so its ask still renders.
+
+## Verification
+
+431 tests (410 → 431). Live: all 21 booths 200, and each of the four verbatim
+booths grew by exactly 46 bytes — `len(EMBED_SCRIPT_TAG)`, one append, nothing
+else. Related: [[2026-09-22-the-browser-became-a-test-surface]],
+[[2026-09-22-seven-of-seven-falsifiers]],
+[[2026-09-21-regex-injecting-chrome]].

persistent-memory.d/2026-09-22-u4-derived-lifetime-landed.md

@@ -0,0 +1,50 @@
+# U4 landed — lifetime is derived, not declared
+
+_2026-09-22 · booth_
+
+**A booth's lifetime stopped being a boolean somebody remembered to press.**
+Three states now, and `sweep_once` is the only thing that honours the first two:
+
+    KEPT       `.forever` present                        never swept   (unchanged)
+    HELD       an open pick, or marks we cannot read     never swept   (new)
+    EPHEMERAL  everything else                           24h           (unchanged)
+
+Plus **viewing is activity**: a deliberately-served response from a booth's own
+page route writes `.viewed`. That dotfile is not a `.lock` dotfile, so
+`_newest_mtime` already counts it — **there is no new arithmetic anywhere**.
+`booth_age_seconds`, `is_expired` and `expires_in` are byte-for-byte what they
+were. A view is one more thing in the tree, which is the same trick `.booth.json`
+used in U5.
+
+**What counts as a view, and why the exclusions matter more than the inclusions.**
+`/b/<n>/` (gallery, verbatim report, `?download=1` zip), `/b/<n>/view` and
+`/b/<n>/marks` count. `/b/<n>/marks.json`, asset GETs, `/`, `/healthz` and a
+zoom URL that 404s do NOT. The marks.json exclusion is load-bearing: **an agent
+must not be able to hold its own booth open by polling for the answer it is
+waiting on.** `/b/<n>/asks` is a 308 into `/marks` and records through it — one
+call, not two.
+
+Checked because it would have been silent: **nothing in the fleet polls a booth
+page.** Homepage's `siteMonitor` for the Booth is `/healthz`, which is on the
+not-a-view list. Had it been pointed at a booth URL, every booth would have
+become immortal on deploy and nothing would have reported it.
+
+**The hold is unbounded and that is the point** — unanswered is unfinished. What
+makes it safe is visibility plus two exits that already existed: the card and
+every Booth-owned header say `held until answered` where the countdown was, and
+`booth rm` / the UI x / `DELETE /b/<n>` take a held booth exactly as they take a
+kept one. **A hold is protection from the timer, never from the operator.**
+
+**Release is activity, stated rather than accidental.** Releasing a kept board
+still buys a full TTL — unchanged — but now because `booth_unkeep` calls
+`record_view`, which is a rule, and no longer because unlinking a file happened
+to bump a directory's mtime, which is not. The CLI warning against
+"unkeep and let it expire" stays and stays true.
+
+⚠ **Running `scripts/layout-probe.py` over booth pages resets every booth's
+clock**, because a GET of a booth page is a view and the probe is not exempt
+from its own rule. Harmless, recoverable, and noted in the probe so nobody
+debugs it later as a sweeper that stopped working.
+
+Contract: `docs/contracts/u4_derived_lifetime.contract.md`. Both heid panels ran
+and the bug hunt after them; see the sibling entries.

persistent-memory.d/2026-09-22-u5-adoption-split-in-two.md

@@ -0,0 +1,23 @@
+# U5's adoption prediction split in two
+
+_2026-09-22 · booth_
+
+**U5's adoption prediction, SPLIT IN TWO within an hour of
+landing — and the split is the interesting part.** The baseline was recorded as
+0 of 26. Fifty minutes after the deploy, `comfy-dev` created `muse-clothed-repro`
+and it announced itself: `{handle: comfy-dev, why: "", created: ...}`. That peer
+was told nothing. **The HANDLE propagates for free** — it rides on `booth new`
+and `booth add`, so every existing CLI caller starts announcing without learning
+anything, which is the flags-on-existing-verbs decision paying off on day zero.
+**The WHY does not** — it needs someone to know the flag exists, and this first
+one is empty.
+
+So re-measure BOTH on **2026-09-29**, because they answer different questions:
+
+    find ~/booth-data -maxdepth 2 -name .booth.json | wc -l          # free
+    grep -l '"why": "[^"]' ~/booth-data/*/.booth.json 2>/dev/null | wc -l   # learned
+
+A high first count and a near-zero second is the predicted shape of "nobody was
+told", and it is the case the operator's no-announcement decision was designed
+to be able to see. Do not read the n=1 above as a rate — it is a code-path
+observation (every CLI caller writes a handle), not a sample.

persistent-memory.d/2026-09-22-u5-panels-reached-a-released-bug.md

@@ -0,0 +1,21 @@
+# Two U5 panels, and prose reached a released outage
+
+_2026-09-22 · booth_
+
+**Two cross-frontier panels on U5, and a paraphrase panel reached
+a production outage two modules away.** 3-of-4 flagged the contract's "4 GB"
+case as letter-compliant but purpose-defeating; the conformance round found that
+unbounded read live in U5's code; walking it to the sibling found the SAME hole
+**live in released `v0.2.2`** — `marks._read_raw` catches `(OSError, ValueError,
+UnicodeDecodeError)` and `json.loads` on deep nesting raises **RecursionError**,
+which is none of them, so 400 KB of brackets in one booth returned 500 for `/`
+and `/healthz` across all 26. The v0.2.2 round HAD flagged it and I closed half:
+**a finding with two call sites is not closed when one is.** The reusable
+instruction — **walk a conformance finding to the sibling module even when the
+sibling is out of scope.** Five of ten conformance findings were tests of mine
+that pass on the regression they exist to catch, three of them asserting an
+ARTIFACT of the property rather than the property; that is three nights running
+on the same shape. Two real bugs neither my tests nor I could see: a bare
+`booth add` wiped the `why` on the one sequence the feature exists for, and
+`--title` was write-only. Full triage in
+`persistent-memory.d/2026-09-22-u5-panels.md`.

persistent-memory.d/2026-09-22-u5-panels.md

@@ -0,0 +1,102 @@
+# U5's two cross-frontier panels — full triage
+
+**Date:** 2026-09-22 · **Paraphrase:** thread `01M340PNVRS21HPASZT38PXQPN` ·
+**Conformance:** thread `01M341E9XAPZEFBSPK9HPGAM0S` · **Shipped as:** `v0.3.0`
+
+Two four-arm artifact-only rounds, dispatched ~30 minutes apart and correctly
+firewalled: the paraphrase ran the **pre-seam-review** capture (073612), the
+conformance round the **SR-amended** one (074901). Heid diffed the two at
+intake and said so.
+
+The conformance round's honest headline is Kimi's: **zero drift in the strict
+sense — the code is a clause-for-clause implementation of the contract.** Both
+rounds' weight landed one layer down, in test strength and contract finish.
+
+## The result worth keeping
+
+**A paraphrase panel reading nothing but prose reached a production outage two
+modules away.** 3-of-4 flagged INV-2's "4 GB" case as *letter-compliant but
+purpose-defeating* — the invariant constrained the RETURN, not the cost, so an
+unbounded read "recreates the outage in slow motion". The conformance round then
+found that exact unbounded read live in U5's shipped code. Walking it to the
+sibling module found the same hole **live in released `v0.2.2`**: `marks.py`'s
+`_read_raw` catches `(OSError, ValueError, UnicodeDecodeError)`, and
+`json.loads` on a deeply nested document raises **RecursionError**, which is
+none of them. A 400 KB file of nothing but brackets in any ONE booth returned
+500 for `/` and `/healthz` across all 26.
+
+**The v0.2.2 round had flagged this and I closed half of it.** Kimi's R5(c)
+named RecursionError explicitly; I adopted "wrap `_hydrate` per-entry" and left
+the `json.loads` above it unguarded. **A finding with two call sites is not
+closed when one is.**
+
+**The reusable instruction: walk a conformance finding to the sibling module
+even when the sibling is formally out of scope.** Heid captured it as its own
+lesson.
+
+## The densest class was tests that could not fail
+
+Five of ten adopted conformance findings were tests of mine that pass on the
+regression they exist to catch. Three shared one shape — **asserting an
+ARTIFACT of the property instead of the property**:
+
+| test | asserted | should have asserted |
+|---|---|---|
+| `test_the_write_is_atomic` | no `*.tmp` survived | the inode changes (`write_text` leaves no temp file either) |
+| INV-3 preservation | a stamp survived a window shorter than the stamp's own resolution | a stamp from 2019 |
+| `test_announcing_is_activity` | age via the directory mtime, which the write bumps either way | the file's own mtime, directory clock restored |
+
+That is the same shape as the marks round's guard-strength finding the night
+before — **three nights running**. Proposed to heid as a standing
+"green-tests-prove-nothing" direction for the skill; routed to the operator
+alongside two other methodology proposals from the same night.
+
+⚠ **My first replacement for the atomicity test was ALSO vacuous.** It spied on
+`os.open` to prove the published path was never written directly — which passes
+trivially, because `Path.write_text` reaches the syscall through `io.open` in C
+and never touches the Python-level `os.open`. The dead end is recorded in the
+test's own docstring rather than deleted.
+
+## Two real bugs the tests were structurally blind to
+
+**`booth new x --why "…"` then `booth add x out/*.png` erased the why.** Omitted
+flags meant empty strings; empty strings overwrote. Two arms predicted it *from
+the contract's wording alone* — "gains a manifest with no `why`" does not
+distinguish a first write from a re-announce with the flags omitted. Every test
+written for this module passed `--why` on both calls, so none could see it.
+Omitted means unchanged now; `--why ""` still clears. The shell carries the
+distinction by leaving the variable UNSET, not empty.
+
+**`--title` was write-only** — stored, flag-surfaced, rendered nowhere. 4-of-4,
+independently top-ranked by every arm of the paraphrase round. It renders on the
+booth page heading with the directory name kept beside it, because the directory
+name is the identity the operator navigates by and refers to positionally.
+
+## Contract-finish, and why it mattered
+
+**INV-1 contradicted its own falsifiable criterion** (4/4) — "the only place
+`.booth.json` is opened" versus INV-3's read-back, which forces `write_manifest`
+to open it. One half was already false of a correct implementation. Restated as
+*one module knows the filename*, which is true, falsifiable and now tested.
+
+**INV-5 named two different promises** (3/4) — the repo's atomic-write rule and
+this unit's render rule. Repo-wide rules are named in words now, never by a bare
+number that can collide with a local one.
+
+Regin's meta-observation is the round's methodology keeper and was borne out:
+**flags cluster where the same rule is re-voiced per signature**, and four of
+eleven contract edits were reconciling a docstring against a prose section
+saying the same thing slightly differently. A table-vs-signature consistency
+pass would beat the format's prose bias.
+
+## Declined / parked
+
+- **Custom booth pages skip provenance** (hulda, solo, verified) — settled
+  independently as U3's seam ~20 minutes before the reply landed. Convergence,
+  not an adoption.
+- **Empty-handle coercion misattributes to the service** — kept, documented. A
+  manifest naming no handle does not read back at all, and an unreadable file is
+  the worse outcome. Unreachable from the CLI.
+- **`used`-set: `touches` versus SR-1 unreconciled** — the code adds the entry
+  as consistency with the equally-unreachable `UPLOAD_MARKER` entry that
+  predates this unit, and says so rather than claiming it prevents anything.

persistent-memory.d/2026-09-22-u6-benches-landed.md

@@ -0,0 +1,67 @@
+# U6 landed — three surfaces, three jobs, one predicate
+
+_2026-09-22 · booth_
+
+**The sixth of seven v1 units. Only U7 is left.** 444 → 555 tests, suite green,
+deployed and verified live: 23/23 booths 200, and the board renders **156 dead
+of 221 rows** — the exact count an independent shell measurement produced before
+a line of code was written, from two different implementations.
+
+## What shipped
+
+- **`booth/benches.py`** (new, stdlib-only AND sibling-free): `Bench`,
+  `normalize_bench_url`, lenient `read_benches`, strict `upsert_bench`,
+  `set_bench_state`, `remove_bench`, `order_benches`. Registry at
+  `~/booth-data/.benches.json` — a dotfile at the DATA ROOT, keyed by id, so two
+  rows with one identity are impossible by construction.
+- **`links.booth_target`** — ONE predicate for "is this a booth URL", consumed
+  by three callers (the CLI refusal, the board's dead marker, `bench import`).
+  Host-agnostic and path-shaped; percent-decodes the name.
+- **`booth link` refuses a booth URL**, names `booth new --why`, and writes
+  nothing — not even the board directory.
+- **The board marks dead rows.** Removal stays the operator's two clicks through
+  the bulk control that already existed. Nothing in the unit deletes a row.
+- **`booth bench add|ls|state|rm|import`**; `import` writes nothing without
+  `--apply` and never touches `links.md`.
+- `docs/archive/links-2026-09-22.md` — the board archived verbatim into git.
+
+## The decision that mattered most, and it was measured
+
+**Identity is the FULL normalized URL, not the origin.** Collapsing the 43
+non-booth rows by origin gives 19 groups; by full URL, 35. The difference is not
+duplication — it is **eight distinct gitea repos merged into one**, three
+unrelated HuggingFace model cards merged into one, and **the two LRPG surfaces
+on `10.100.10.50:8321`, which are the IA doc's own example of two real benches**,
+merged into one. Origin identity destroys more than it dedups. Full-URL identity
+still collapses both cases the doc names (talk 5→1, Peedlar 3→1).
+
+Query is IN the identity (three ShutterChute rows differ only by `?token=` and
+are three real links); fragment is OUT; credentials are REFUSED, not stripped.
+
+## The seam review earned it again — three real contract defects
+
+Run in-session against the real `.py` files, after the cold panel was dispatched:
+
+- **SR-1** — the contract claimed `test_stdlib_only` already forbids sibling
+  imports. **It does not**: its failure set is `{r for r in roots if r !=
+  "booth" and ...}`, which exempts `booth` on purpose. Only test_manifest.py has
+  the strict copy. INV-9 would have shipped untested.
+- **SR-2** — the contract named `resolve_booth` as the dead marker's existence
+  check. That function is a closure inside `create_app` and **raises
+  HTTPException(404)** — per row, one swept booth would 404 the whole board page.
+- **SR-7** — booth links are emitted through `quote(name, safe="")`, so a
+  predicate comparing the raw segment marks every encoded-name booth dead
+  forever.
+
+SR-4 and SR-5 were **verified rather than assumed**: both `list_booths` and
+`sweep_once` skip a child that is not a directory AND one whose name starts with
+a dot, so the registry is safe from the sweeper by two guards, not one. Had
+either been absent the design would have eaten its own registry on tick one.
+
+## How it closed
+
+All three cold gates came back and were folded in full, with exactly one finding
+declined. Released as `v0.6.0` — see [[2026-09-22-u6-benches-released]] and
+[[2026-09-22-three-cold-panels-on-one-unit]]. The tag waited for the gates, per
+the v0.2.0 lesson, and that sequencing was right: the panels produced ten code
+fixes after this entry was first written.

persistent-memory.d/2026-09-22-u6-benches-released.md

@@ -0,0 +1,58 @@
+# U6 released as v0.6.0 — benches, and the number that was two defects
+
+_2026-09-22 · booth_
+
+**The sixth of seven v1 units. Only U7 remains.** 444 → 607 tests. Tagged
+`v0.6.0` (minor, operator-approved). **NOT PUSHED** — push is his call.
+
+## What shipped
+
+- **`booth/benches.py`** — stdlib-only AND sibling-free. `Bench`,
+  `normalize_bench_url` (the identity), a lenient `read_benches` on the render
+  path and a strict `_load_strict` on the write path, `mkstemp` + `fsync` +
+  `os.replace` under an flock, and `order_benches` with a stated total order
+  `(state rank, name casefolded, id)`.
+- **`links.booth_target`** — ONE predicate for "is this a booth URL",
+  host-agnostic, path-shaped, percent-decoding, control-character-rejecting,
+  never raising. Three callers: the CLI refusal, the board's dead marker,
+  `bench import`.
+- **`booth link` refuses** a booth URL (naming `booth new --why`) and a
+  credentialed one, writing nothing in either case.
+- **The board marks dead rows** — 161 of 221 live. Removal stays the operator's
+  two clicks through the bulk control that already existed. Nothing deletes.
+- **`booth bench add|ls|state|rm|import`**. `--apply` REQUIRES the ids.
+
+## The decision that shaped the unit, and it was measured
+
+**The design doc's headline "69% rot" was two defects wearing one number**, and
+splitting them is what made the unit the right size — see
+[[2026-09-22-one-number-was-two-defects]]. 178 of 221 rows are booth
+announcements (156 already dead) whose *cause* U5 had already closed; only 8 are
+the bench re-post the registry fixes. A unit scoped off the unsplit number would
+have built the registry, declared victory, and left 178 rows rotting.
+
+**Identity is the FULL normalized URL, not the origin**, and that was measured
+rather than chosen: origin identity merges eight distinct gitea repositories
+into one row, three unrelated HuggingFace model cards into one, and the two LRPG
+surfaces on `10.100.10.50:8321` — *the design doc's own example of two real
+benches* — into one. It destroys more than it deduplicates.
+
+**`booth link` is NOT deprecated**, against the design doc's plan. Roughly 14 of
+the 35 distinct non-booth targets are reference bookmarks (repos, model cards,
+docs) for which the board is the right and only home. Deprecating it would have
+evicted a third of its live content. The IA doc is corrected.
+
+## The gates
+
+All four closed, and every one paid — see
+[[2026-09-22-three-cold-panels-on-one-unit]]. Contract review
+`01M35BWCJ806MT75NA630Y4WFH`, code review `01M35CK8YKEKMV7T15JXEF6A8N`, bug hunt
+`01M35CRRK2RTVWWF1BN09AFQG3`, one consolidated reply sent to heid at
+`01M35FY8QZTB9E5VR4WXDSBGEV`.
+
+## Live evidence, unplanned
+
+The sweeper ran mid-session: **23 booths → 19**, and dead board rows went
+**156 → 161 in about fifteen minutes**. The defect compounding in real time
+while the fix was being built — which is the argument for U6-before-U7 playing
+out on its own.

persistent-memory.d/2026-09-22-u7-landed-and-a-table-that-did-not-reproduce.md

@@ -0,0 +1,77 @@
+# U7 landed — and the number that justified it did not reproduce
+
+_2026-09-22 · booth_
+
+**The last v1 unit is in.** The three ratified components landed at `a306e2d`;
+the fourth — filename-prefix groups replacing subfolder sections — landed here,
+with `test_no_group_rail_is_shipped_yet` deleted in the same commit that built
+what it guarded against. **All seven v1 capabilities are now landed.**
+
+## The part worth remembering: the contract's own measurement was wrong
+
+The contract stated a rule and, beside it, a table of what that rule produced.
+**They are not the same computation.** Implementing the stated rule and running
+it against the live set:
+
+| booth | contract claimed | stated rule actually gives |
+|---|---|---|
+| `sindra-corpus-v1` | 16 | 16 ✓ |
+| `sindra-sfw-pool` | 10 | 10 ✓ |
+| `sindra-nude-pool` | 12 | 12 ✓ |
+| **`sindra-bakeoff`** | **5** | **24** |
+| **`sindra`** | **1 (degenerate)** | **27** |
+
+Three of five matched, which is what made it survive review. The two that did
+not were **the two load-bearing rows**: bakeoff was the "this pays" evidence and
+sindra was the degenerate case INV-3 was written for.
+
+**The contract contradicts itself in plain sight and nobody caught it.** Its own
+worked example says `00-sheet-c1-market-noon.png` has no trailing digit run and
+therefore groups as its whole stem — which makes eight of bakeoff's forty images
+eight singleton groups, so 5 was never reachable. And the numbers ARE
+reproducible, just not by one rule: **first-two-segments gives exactly 5 on
+bakeoff; first-segment gives exactly 1 on sindra.** The table was assembled from
+two different heuristics and written up as one.
+
+⚠ **A cold contract-review panel cannot catch this, and did not.** The panel
+reads the artifact; the artifact is internally plausible. Only running the
+stated rule against the live data falsifies it. **A measurement inside a
+contract is not reviewed by reviewing the contract** — it is reviewed by
+re-running it, and that is now a thing to do before implementing any contract
+whose scope rests on a number.
+
+## The degeneracy it guarded was the wrong one
+
+INV-3 guarded **one group for everything** ("a rail with one entry cannot
+navigate"). The live set's actual failure is the opposite: **one group per
+item** — `pewpew-ui-brief` 23 groups for 34 items, `dfa-concepts` 13 for 20. The
+contract as written would have shipped a 23-row rail that is a second copy of
+the grid. INV-3 now guards both, with a live specimen each:
+
+- **(a)** `sc-iso-spread` — `DSC0001.jpg`–`DSC0006.jpg`, one group of six.
+- **(b)** `pewpew-ui-brief` — 23 groups, 19 of them singletons.
+
+The shipped predicate, one line: **two or more groups, and the middle group
+holding more than one item.** It gets all 17 booths right.
+
+## The shipped rule, and why it differs
+
+`strip ONE trailing run of digits` keys on the END of the stem, which is where
+the *instance number* lives — so it splits `m-c1-market-noon-9401` from
+`m-c2-rain-street-9403`, which are the same family. The shipped rule keys on the
+**first separator-delimited segment**, where the family lives, destemming only
+when the stem has no separator at all (so `ac01` → `ac`, but `v30-seed8302` and
+`v35-seed8302` stay apart — that split is the axis `muse-clothed-repro` is
+about).
+
+Live result: `sindra-corpus-v1` renders `ac 12 · bu 10 · cu 12 · fb 12 · … ·
+wu 8` over 66 images. `sindra-bakeoff` renders `00 · README · m · r`, which are
+its three real families.
+
+## Also true, and easy to trip on
+
+**`miranda-is` and `sindra-voice-1` group beautifully and get no rail** — both
+carry `index.html`, so they take the verbatim path and have no grid at all. A
+measurement taken with `booth_items` alone predicts a rail for them; the route
+does not. Measure the RENDERED surface, not the resolver, when the question is
+"what will the operator see".

persistent-memory.d/2026-09-22-u7-remeasured-before-scoping.md

@@ -0,0 +1,70 @@
+# U7 re-measured before scoping — sections are dead, filename prefixes are not
+
+_2026-09-22 · booth_
+
+**Pre-work, not the unit.** The standing instruction is "re-count the booths
+before scoping U7". Done, on the live set (19 booths). No U7 code, no U7
+contract — this exists so the scope call is a thirty-second read.
+
+## The set as it actually is
+
+| booth | items | images | subdirs | shape |
+|---|---|---|---|---|
+| `miranda-is` | 92 | 0 | 0 | report |
+| `sindra-bakeoff` | 81 | 40 | **0** | gallery |
+| `sindra-corpus-v1` | 66 | 66 | **0** | gallery |
+| `sindra` | 61 | 30 | **0** | gallery |
+| `sindra-sfw-pool` | 59 | 59 | **0** | gallery |
+| `sindra-nude-pool` | 42 | 42 | **0** | gallery |
+| `pewpew-ui-brief` | 34 | 1 | 7 | **report** |
+| `dfa-concepts` | 21 | 14 | 1 | **report** |
+| …11 more | ≤19 | | 0 | |
+
+## Finding 1 — sections are worth ZERO, and this is now measured twice
+
+**Not one gallery booth has a subdirectory.** Zero of eleven. The only two
+booths with subfolders are both **reports**, the job where grid navigation
+matters least, and `pewpew-ui-brief`'s seven subdirs hold one image.
+
+The IA doc calls sections "most of the navigation fix". On this set they are
+none of it. Cutting `Item.section` rendering from U7 costs nothing measurable.
+(`Item.section` already exists from U1 and stays — this is about whether U7
+builds a section RAIL, not about deleting a field.)
+
+## Finding 2 — the grouping signal is in the FILENAME, and it pays
+
+Tested two heuristics against every large gallery. Strip a trailing digit-run
+from the stem and group on what remains:
+
+| booth | images | groups | verdict |
+|---|---|---|---|
+| `sindra-corpus-v1` | 66 | **16** | useful |
+| `sindra-nude-pool` | 42 | **12** | useful |
+| `sindra-sfw-pool` | 59 | **10** | useful |
+| `sindra-bakeoff` | 40 | **5** | useful |
+| `sindra` | 30 | **1** | **degenerates** |
+
+Specimens: `00-sheet-c1-market-noon.png`, `ac01.png`, `a01.png`,
+`flag-rear.png`. The competing heuristic — split on the second hyphen — is
+useless everywhere (59 "groups" from 59 files).
+
+So a prefix heuristic pays on **4 of 5** large galleries and collapses to one
+group on the fifth. **That is a filter/grouping affordance, not a section
+rail**, and it must degrade gracefully to "one group" rather than render a
+useless single-section rail.
+
+## What this implies for the scope, stated as a recommendation not a decision
+
+U7 as written is four things: sections, a sticky rail, filters, grid keyboard.
+The measurement says **drop sections, keep the other three**, and consider
+prefix-grouping as the thing sections were supposed to be — with a stated
+degenerate case.
+
+⚠ The sizing case has also changed: the unit was scoped against 270-item
+booths and **the largest gallery is now 81 items / 40 images**. Everything
+about virtualization stays parked ([[2026-09-21-ia-and-v1-gate-landed]] names
+it); at 66 images a lazy grid is fine and measuring it first is the rule.
+
+**The booth set churned again during this session** — `sindra-sfw-pool` (59
+images) appeared and the `pancake-*` set went. Re-count again before writing
+the contract; do not trust this table either.

persistent-memory.d/2026-09-22-u7-three-quarters-and-one-ruling.md

@@ -0,0 +1,56 @@
+# U7 is three-quarters built and blocked on one word
+
+_2026-09-22 · booth_
+
+**The last v1 unit, decomposed by what the operator has already ratified versus
+what he has not.** ROADMAP's U7 row names four components. Three were already
+approved there and are **built, tested and deployed** (`a306e2d`). The fourth is
+a scope departure and is **deliberately not built**.
+
+| component | ROADMAP | state |
+|---|---|---|
+| sticky rail | ratified | **landed** — totals + per-filter counts |
+| filters | ratified | **landed** — all / flagged / annotated / unanswered |
+| grid keyboard | ratified | **landed** — `←/→ f n Enter Esc`, bound only when a grid exists |
+| **sections → filename groups** | **departs** | **NOT BUILT** |
+
+`tests/test_navigation.py::test_no_group_rail_is_shipped_yet` fails the moment
+somebody builds the group rail anyway, so the departure cannot arrive by
+accident while the ruling is outstanding.
+
+## The question, and why it is his
+
+**Drop subfolder sections for filename-prefix groups — yes or no?**
+
+Measured (see [[2026-09-22-u7-remeasured-before-scoping]]): **zero of eleven
+gallery booths have a subdirectory**, so sections buy nothing; stripping a
+trailing digit-run from the stem yields **5–16 sensible groups on four of the
+five large galleries** and degenerates to one group on the fifth. The
+replacement is better on the evidence — but swapping a ratified component for
+an unratified one is scope direction, not implementation.
+
+Contract at `docs/contracts/u7_navigation.contract.md`, status
+`PARTIALLY LANDED`, with the departure named as the operator's call.
+
+## Decisions taken under stated assumption, both cheap to reverse
+
+- **`unanswered` means HAS AN OPEN PICK** — the U4 hold predicate, which already
+  exists. The other reading ("has no mark at all") is a genuinely different
+  question and stays an open question on the contract.
+- **Filters are LINKS, not scripts**, resolved server-side, so the gallery keeps
+  working with JavaScript off. U3 cost the verbatim path its no-JS operation and
+  said so plainly; the gallery is the surface the operator actually reviews on,
+  and this unit does not repeat it there.
+
+## The vacuous falsifier, written an hour after the entry about them
+
+`test_filtering_never_reorders` compared each filtered view against the
+**unfiltered response** — so a mutation reversing the order reversed both sides
+and it **stayed green under the exact change it forbade.** Caught only by
+running the mutation rather than trusting the assertion.
+
+Rewritten against an independent truth: U1 INV-3 says the order IS `sorted(rel)`,
+so each view must be sorted, full stop, with no reference to another response.
+Re-verified RED. **Every new falsifier in this session was mutation-checked
+after this**, and that is the practice to keep — see
+[[2026-09-22-vacuous-falsifiers]].

persistent-memory.d/2026-09-22-vacuous-falsifiers.md

@@ -0,0 +1,40 @@
+# Five of seven INV falsifiers did not falsify anything
+
+_2026-09-22 · booth_
+
+The U4 contract carried seven invariants, each with a *Falsifiable:* line, and
+each had a test. **The code-review panel showed that five of the seven tests
+would still pass under a change that defeats the invariant they name.** Gróa's
+"per INV entry, what would still pass" section is the single most useful thing
+either panel produced on this unit.
+
+| INV | what the test asserted | what still passed |
+|---|---|---|
+| 1 (no new arithmetic) | the clock moved after a view | special-casing `.viewed` inside `_newest_mtime` — the exact new arithmetic INV-1 forbids |
+| 3 (`is_held` is pure) | the right answer, once | `is_held` doing I/O, or `return True` unconditionally |
+| 4 (every surface says why) | a substring on `GET /` | dropping the line from the booth header, the marks page, or the board branch |
+| 5 (a view cannot fail a request) | `record_view` did not raise | a second `touch` outside the guard, 500ing all three routes |
+| 6 (unreadable marks hold) | the corrupt booth survived | a sweeper that deletes nothing at all (no doomed sibling in the fixture) |
+| 7 (machine reads do not hold) | `.viewed` was absent | a handler writing any other non-dot file, holding the booth open just as well |
+
+**The shape of the error is the same every time: the test asserted the OUTCOME
+the author was thinking about, not the DISCRIMINATOR the invariant names.** A
+green test proved the happy path and nothing about the invariant. Writing the
+falsifiable line in the contract did not produce a falsifying test — it produced
+a test that *cited* one.
+
+Fixed by rewriting each to fail under the change that defeats it: same-mtime
+equivalence with an arbitrary non-lock dotfile (plus a `.lock` that must NOT
+count); `is_held` called with marks belonging to a booth that does not exist on
+disk; one test per rendered surface, each rendering only its own; the three
+routes GET against a chmod'd booth; a doomed sibling; the AGE asserted rather
+than the marker. **The board-header pair was verified RED against the pre-fix
+template rather than assumed** — which is the step that makes "fixed, not
+amended" trustworthy.
+
+**The method to keep: for each invariant, name a change that defeats it and ask
+whether the test goes red.** If you cannot name one, the invariant is not
+falsifiable yet. Regin and Kimi independently proposed this as a contract-time
+"vacuity pass"; heid rates this round the strongest evidence for it so far, and
+it is a `/heid*` skill proposal sitting with the operator, not a change to this
+repo.

persistent-memory.md

@@ -1,6 +1,6 @@
 # Persistent memory — booth
 
-_Last updated: 2026-09-21_
+_Last updated: 2026-09-22_
 
 > **Always check for `/tmp/booth-dev-handoff.md`** — if it exists and its
 > `Written:` stamp is under 8 hours old, read it (it carries the in-flight
@@ -17,190 +17,125 @@ loop it turned out to actually be.
 
 ## Current state / in-flight
 
-_As of 2026-09-21:_
+_As of 2026-09-22:_
 
-- **v1 is gated on seven units** in `ROADMAP.md`, ordered by dependency:
+- **U6 SHIPPED (`v0.6.0`) with a late fix (`v0.6.1`). PUSHED.** `main` and both
-  **U1 → U2 → {U3, U4, U5} → U7**, with **U6 independent** of all of them.
+  tags are on `origin` as of 2026-09-22 — the tree is no longer single-copy.
-- **U1 (one item record) has landed** at `ce598b3` and is verified against its
+  → `persistent-memory.d/2026-09-22-u6-benches-released.md`
-  own invariants, not just its commit message: INV-1 holds (no `classify` /
+- **THE OPERATOR RULED ON EVERYTHING OUTSTANDING (2026-09-22, "accept all
-  `doc_kind` / `read_blurred` / `render_doc` call survives in a route body),
+  recs").** Four of five settled and executed; one blocked by the permission
-  the zoom and doc templates render the caption they now receive, the
+  layer. Nothing is waiting on him. →
-  re-exports are asserted by a test. 192 tests green, `0.1.15`.
+  `persistent-memory.d/2026-09-22-operator-ruled-on-the-open-five.md`
-- **U2 (marks) has landed** — `booth/marks.py`, contract at
+- ✅ **U7 IS LANDED — ALL SEVEN v1 UNITS ARE IN.** The fourth component
-  `docs/contracts/u2_marks.contract.md`, 242 tests green. Not yet deployed.
+  (filename-prefix groups) is built; `test_no_group_rail_is_shipped_yet` was
-- **U2 is DEPLOYED and the migration is done.** The service was restarted
+  deleted in the same commit, as required.
-  2026-09-21 23:41 and again after the `auto_reload` fix; all four legacy
+  → `persistent-memory.d/2026-09-22-u7-landed-and-a-table-that-did-not-reproduce.md`
-  sidecars imported (`dfa-concepts/dfa`, `run07-decisions/decisions`,
+- 🔶 **THE 1.0 CUT IS NOW A DECISION, NOT A DEPENDENCY, AND IT IS HIS.** The v1
-  `sc-iso-spread/spread`, `sindra-voice-1/anchor`, all still open) with the
+  target is met. A major bump needs explicit operator approval; nothing in the
-  sidecars left on disk. Verified live: index + 25 booths x {booth page, marks
+  code is waiting on it. The open fork: cut `1.0`, or stage a `0.7.0` first.
-  page, marks.json} all 200, plus zoom views on five booths.
+  **Not bumped — the work is committed as commits, which are not releases.**
-- **Still needs the operator: the release tier.** U2 changes the CLI surface for
+- ⚠ **U7'S CONTRACT CARRIED A MEASUREMENT THAT DID NOT REPRODUCE**, and it was
-  17 consuming handles (`booth asks` -> `booth marks`, new `marks-import`) and is
+  the number the scope departure rested on. The stated rule gives 24 and 27
-  a v1 unit, so it reads minor-worthy — which needs explicit approval per the
+  groups where the table claimed 5 and 1; the table was assembled from two
-  SemVer rule. Nothing is bumped or tagged; the work is committed as SHAs.
+  different heuristics. **A cold contract-review panel cannot catch this** — the
-- **`/heid-contract-review` on the U2 contract is still in flight** (panel mode,
+  artifact is internally plausible. Re-run any measurement a contract's scope
-  posted 2026-09-21, redacted copy at
+  rests on before implementing it. Same detail file.
-  `/tmp/heid-contract-review/booth-20260922-061015/`). Triage it when it lands —
+- ⚠ **A MUTATION HARNESS NEEDS A GREEN BASELINE AND CACHE DEFEAT**, or it
-  the code is written, so findings land as follow-up fixes rather than contract
+  certifies falsifiers without running them. Both defects bit in one session.
-  edits. The seam review ran in-session and its nine findings are already folded
+  → `persistent-memory.d/2026-09-22-a-mutation-harness-that-certified-a-broken-test.md`
-  into the contract and the code.
+- **639 tests green; 12/12 new falsifiers mutation-proved. Deployed; 21/21
-- **Open, operator's call:** whether U6 (benches) runs in parallel with U2 or
+  booths 200.** ⚠ The set churned again mid-session (19 → 21).
-  strictly after it. Nothing blocks on the answer; U6 touches different storage
+- 🔶 **A bug-hunt panel is IN FLIGHT** — heid thread `01M368G2Y0JMTJ2T7M3JMTXV5Z`,
-  and a different surface, so it cannot be broken by U2.
+  dispatched 2026-09-22 21:31 PDT over the U7-groups diff. If its reply has not
-- Live service is `active` on `:8090` (systemd `--user`), 25 booths.
+  been consumed, drain `/althing:inbox` and triage before treating U7 as closed.
+- 🛑 **STANDING RULING — NO ANNOUNCEMENTS OUT OF THIS REPO, AND THE OPERATOR
+  SENDS THE EVENTUAL ONE HIMSELF** (operator, 2026-09-22). Verbatim: *"no
+  announcements until the entire arc is done, and even then i'll do it myself."*
+  Two clauses, both binding: **(a)** no althing announcement of any kind ships
+  from booth-dev until the v1 arc is COMPLETE — not per-unit, not at the 1.0
+  tag, not "just the peers who consume it"; **(b)** when the arc IS done, the
+  announcement is HIS to send, not a thing to ask permission for. This is
+  STRICTER than `~/.claude/CLAUDE.md`'s broadcast gate, which merely requires
+  approval — here the send is not the agent's to make at all, so *asking* is
+  also out of scope. Do not offer, draft-and-await, or surface it as a pending
+  decision; it is settled and not a standing question.
+  **The 17-handle `booth link` note is consequently REASSIGNED, not blocked.**
+  The full draft + recipient list stays at
+  `docs/pending/fleet-note-booth-link-refusal.md` as MATERIAL FOR HIM. It is no
+  longer an open loop, no longer awaiting approval, and no longer a thing to
+  raise. Same for anything U4's `keep`-semantics change would have warranted
+  telling peers.
+- ⚠ **NOT SEEDED, and this survives the blanket ruling.** "No seeding yet" was a
+  SPECIFIC prior instruction, not a recommendation of mine, so "accept all recs"
+  does not override it. `.benches.json` does not exist in `~/booth-data`.
+- **A U7 directive (D-0011) misrouted to infra-ops and is SUPERSEDED.** Miranda
+  confirmed directly. Nothing to act on.
+  → `persistent-memory.d/2026-09-22-a-directive-misrouted-by-pane-title.md`
+- ⚠ **THE 17 CONSUMING HANDLES WERE NEVER TOLD that `keep` stopped meaning
+  "waiting on an answer"** — and the note above does not tell them either; it is
+  about `booth link`. **This CHANGES HOW THE 2026-10-06 RE-COUNT READS**: a flat
+  `.forever` rate does NOT falsify the diagnosis.
+- **Two dated predictions pending, not to be run early.** U5's adoption
+  re-measure **2026-09-29**; the `.forever` re-count **on or after 2026-10-06**.
+- **FIVE `/heid*` methodology proposals sit with the operator**, untracked by
+  his choice.
+- The booth set churns hard: 26 → 24 → 25 → 23 → **19**. Re-count rather than
+  trusting any number here.
 
 ## Recent decisions
 
-- `[2026-09-21]` **Deterministic order is a cross-cutting v1 invariant** —
+- `[2026-09-22]` ✅ **U7 landed — and the number that justified it did not reproduce** — all seven v1 units are in; READ BEFORE TRUSTING A MEASUREMENT INSIDE A CONTRACT, and before assuming a degeneracy guard covers the degeneracy you actually have → `persistent-memory.d/2026-09-22-u7-landed-and-a-table-that-did-not-reproduce.md`
-  operator directive, mid-implementation. Every ordered collection the Booth
+- `[2026-09-22]` ⚠ **A mutation harness certified a broken test, twice, for two reasons** — no green baseline, and the pyc cache silently reverting same-size mutations; READ BEFORE WRITING ONE → `persistent-memory.d/2026-09-22-a-mutation-harness-that-certified-a-broken-test.md`
-  renders must have a *stated* rule producing the same sequence on every render
+- `[2026-09-22]` 🛑 **STANDING: no announcements out of this repo until the arc is done, and he sends that one himself** — verbatim *"no announcements until the entire arc is done, and even then i'll do it myself."* Stricter than the house broadcast gate: the send is not the agent's to make, so **asking is also out of scope**. The drafted 17-handle note is REASSIGNED to him, not blocked — see the in-flight row above; do not raise it again.
-  of the same state; the rule can be anything defensible (byte order, time, an
+- `[2026-09-22]` **The operator ruled on all five open items at once** — four executed incl. the first push; the broadcast was blocked by the permission layer and is drafted at `docs/pending/` → `persistent-memory.d/2026-09-22-operator-ruled-on-the-open-five.md`
-  explicit number, an arbitrary-but-recorded sequence), but no rule at all is
+- `[2026-09-22]` **U7 is three-quarters built and blocked on one word** — the ratified three landed; the sections-vs-groups departure is NOT built and is the operator's call, tracked at `docs/contracts/u7_navigation.contract.md` → `persistent-memory.d/2026-09-22-u7-three-quarters-and-one-ruling.md`
-  forbidden. It binds harder here than elsewhere because the Booth's job is
+- `[2026-09-22]` **An approved directive misrouted because pane_find addresses by a rolling pane title** — resolved; the MECHANISM is the durable part, reported to infra-ops, untracked by booth-dev → `persistent-memory.d/2026-09-22-a-directive-misrouted-by-pane-title.md`
-  **comparison** — the operator judges tile 47 against tile 47 and refers to
+- `[2026-09-22]` **U7 re-measured before scoping — sections are dead, filename prefixes are not** — PRE-WORK ONLY, no unit started; read before writing U7's contract → `persistent-memory.d/2026-09-22-u7-remeasured-before-scoping.md`
-  artifacts positionally, so an order that moves between renders misfiles a flag
+- `[2026-09-22]` **The last open defect closed, and building its falsifier found another** — the wrong-shaped answer fixed at `_hydrate`; `_safe_fragments` lost its natural trigger and its handler could not survive the failure it handled → `persistent-memory.d/2026-09-22-a-wrong-shaped-answer-500s-the-gallery.md`
-  or a note rather than crashing. Recorded as `ROADMAP.md` § "Cross-cutting
+- `[2026-09-22]` **U6 released as `v0.6.0` — benches, and the number that was two defects** — six of seven v1 units landed, NOT PUSHED → `persistent-memory.d/2026-09-22-u6-benches-released.md`
-  invariant" (with the per-collection table) and `CLAUDE.md` invariant 6, and
+- `[2026-09-22]` **Three cold panels on one unit, and what each lens could only see alone** — READ BEFORE DECIDING TO SKIP A GATE; all five passes found something the others structurally could not → `persistent-memory.d/2026-09-22-three-cold-panels-on-one-unit.md`
-  tested. Still undecided and must be settled before those units ship: **U7's
+- `[2026-09-22]` **U6 landed — three surfaces, three jobs, one predicate** — the seam review caught three real contract defects incl. a per-row `resolve_booth` that would have 404'd the board → `persistent-memory.d/2026-09-22-u6-benches-landed.md`
-  section ordering and compare pairing**, and **U6's bench listing**.
+- `[2026-09-22]` **The 69% link-board rot was two defects wearing one number** — READ BEFORE SCOPING ANY LINK-BOARD WORK; U5 closed the larger half and full-URL-vs-origin identity is a measured call → `persistent-memory.d/2026-09-22-one-number-was-two-defects.md`
-- `[2026-09-21]` **U2 (marks) landed.** One primitive replacing three
+- `[2026-09-22]` **U3 landed — the page declares the seam, the Booth mounts into it** — ten regexes against author HTML replaced by a substring test and a `+` → `persistent-memory.d/2026-09-22-u3-declared-embed-seam-landed.md`
-  mechanisms. `pick` / `note` / `flag` in one `.marks.json` per booth, one read
+- `[2026-09-22]` **A wrong-shaped answer 500s the gallery and the marks page** — PRE-EXISTING (measured at `42ea67f`), NOT U3; the v0.2.2 lesson is only half-implemented → `persistent-memory.d/2026-09-22-a-wrong-shaped-answer-500s-the-gallery.md`
-  path (`marks_for`), one openness predicate (`open_marks`), rendered beside the
+- `[2026-09-22]` **The browser became a test surface** — READ BEFORE TOUCHING `playwright` IN pyproject; the pinned upper bound is the foot-gun, and these tests SKIP rather than fail → `persistent-memory.d/2026-09-22-the-browser-became-a-test-surface.md`
-  artifact on the tile, at full size in the zoom, and in the panel. `flag` and
+- `[2026-09-22]` **A vacuity pass that tries the contract's own mutation agrees with itself** — U3 ran one, reported 7/7, and a cold panel then showed one of the seven was vacuous; READ BEFORE WRITING A *Falsifiable:* LINE → `persistent-memory.d/2026-09-22-seven-of-seven-falsifiers.md`
-  `note` had no write path at all before this — the selection loop
+- `[2026-09-22]` **U4 landed — lifetime is derived, not declared** — three states, viewing is activity, and no new arithmetic anywhere → `persistent-memory.d/2026-09-22-u4-derived-lifetime-landed.md`
-  (`golden-candidates`, `sindra-finalists`, the `pancake-*` ladders) was running
+- `[2026-09-22]` **The `.forever` diagnosis got a live positive control** — 3 of the 4 booths awaiting an answer were ALSO hand-pinned — RE-COUNT 2026-10-06 → `persistent-memory.d/2026-09-22-forever-had-a-live-positive-control.md`
-  through chat. 242 tests. Details worth carrying: `asks.py` kept `normalize_ask`
+- `[2026-09-22]` **No fleetwide notice for U4, and what that does to the prediction** — READ BEFORE THE 2026-10-06 RE-COUNT; a flat rate does not falsify the diagnosis → `persistent-memory.d/2026-09-22-no-notice-and-what-it-does-to-the-prediction.md`
-  and gained `build_answer` (the 2026-09-09 partial-answer semantics preserved by
+- `[2026-09-22]` **Four independent paths to one fail-open delete** — the bug-hunt panel's class, and the zsh word-splitting trap that shipped an empty bundle → `persistent-memory.d/2026-09-22-four-paths-to-one-fail-open-delete.md`
-  moving, not rewriting) and LOST its five sidecar-storage functions;
+- `[2026-09-22]` **Two reads of one file are not one read of one state** — a TOCTOU seam that composes two correct readers into a fail-open delete → `persistent-memory.d/2026-09-22-two-reads-are-not-one-state.md`
-  `GET /b/<n>/marks.json` was added because remote sessions polled
+- `[2026-09-22]` **Five of seven INV falsifiers did not falsify anything** — read before writing a *Falsifiable:* line; a green test cited one rather than being one → `persistent-memory.d/2026-09-22-vacuous-falsifiers.md`
-  `<stem>.answer.json` over HTTP and the sidecar's removal would have taken that
+- `[2026-09-22]` **The third one-branch template miss** — this repo's recurring blind spot; read before adding a fact to any template → `persistent-memory.d/2026-09-22-third-one-branch-template-miss.md`
-  capability with it; `/b/<n>/asks` 308s to `/marks`.
+- `[2026-09-22]` **The size cap opened a service-wide hang** — a FIFO has st_size 0; a bound that trusts it inherits what it does not mean → `persistent-memory.d/2026-09-22-size-cap-opened-a-hang.md`
-- `[2026-09-21]` **A partially-answered pick now counts as OPEN** — declared, not
+- `[2026-09-22]` **An existing test stopped me retiring documented behaviour** — the clean fix for the mtime race would have silently changed TTL doctrine → `persistent-memory.d/2026-09-22-doctrine-not-defect.md`
-  smuggled. The old index badge tested `answer is None`, so a half-answered
+- `[2026-09-22]` **Two U5 panels, and prose reached a released outage** — read the detail before assuming a conformance finding stops at its own module → `persistent-memory.d/2026-09-22-u5-panels-reached-a-released-bug.md`
-  four-question ask read as closed on the index while the panel beside it
+- `[2026-09-22]` **U5's adoption prediction split in two** — the handle rides for free, the why must be learned — RE-MEASURE 2026-09-29 → `persistent-memory.d/2026-09-22-u5-adoption-split-in-two.md`
-  rendered `◐ partial`: the two disagreed about the same booth. Open is the
+- `[2026-09-22]` **The U2 bug-hunt panel was not ceremony** — the lock-unlink race and the TTL guard that was failing at its own job → `persistent-memory.d/2026-09-22-u2-bug-hunt-panel.md`
-  reading that makes U4 correct — a lifetime rule that unpinned a booth on the
+- `[2026-09-22]` **The lenient reader's blast radius was the whole service** — marks_for runs per booth per index load; a raise there is an outage → `persistent-memory.d/2026-09-22-lenient-reader-blast-radius.md`
-  first radio click would sweep a review in flight.
+- `[2026-09-22]` **`booth marks` / `booth answer` got real exit codes** — read it before changing anything the 17 consuming handles call → `persistent-memory.d/2026-09-22-cli-exit-codes.md`
-- `[2026-09-21]` **The U2 seam review earned its place, and the record should
+- `[2026-09-22]` **`scripts/booth` went from zero tests to five** — they run the real script under system python3, so they also check INV-1 → `persistent-memory.d/2026-09-22-scripts-booth-got-tests.md`
-  say how.** Nine findings against the real `booth.asks` / `booth.items` /
+- `[2026-09-21]` **v0.2.0 was tagged while a gate was in flight** — the sequencing lesson: if a gate is outstanding, the tag waits → `persistent-memory.d/2026-09-21-v020-tagged-with-a-gate-in-flight.md`
-  `booth.inline` surfaces, two of which changed scope or behaviour: `inline.py`
+- `[2026-09-21]` **A write over a damaged `.marks.json` wiped the booth** — the reads-lenient / writes-strict asymmetry, and why it exists → `persistent-memory.d/2026-09-21-marks-write-wiped-judgment.md`
-  was missing from `touches` entirely (its `place()` indexes asks by
+- `[2026-09-21]` **Seam review and cold panel had zero overlap, twice** — evidence for running both; neither substitutes for the other → `persistent-memory.d/2026-09-21-two-gates-are-complementary.md`
-  **subscript**, which a frozen dataclass refuses — nothing else in the service
+- `[2026-09-21]` **Every code-changing finding came from the AMBIGUITY pass** — a finding about the /heid-contract-review skill, not about this repo → `persistent-memory.d/2026-09-21-ambiguity-pass-did-the-work.md`
-  does that), and the partial-answer inconsistency above. The cold
+- `[2026-09-21]` **Deterministic order is a cross-cutting v1 invariant** — operator directive; read before adding ANY ordered surface → `persistent-memory.d/2026-09-21-deterministic-order-invariant.md`
-  `/heid-contract-review` pass is artifact-only by design and structurally
+- `[2026-09-21]` **U2 (marks) landed — one primitive for three mechanisms** — what moved where, and the HTTP mirror remote sessions poll → `persistent-memory.d/2026-09-21-u2-marks-landed.md`
-  cannot see a sibling module, so neither it nor a same-model self-review would
+- `[2026-09-21]` **A partially-answered pick counts as OPEN** — declared, not smuggled; it is the reading that makes U4 correct → `persistent-memory.d/2026-09-21-partial-answer-counts-as-open.md`
-  have found either. Two more surfaced later and are worth the same note: a
+- `[2026-09-21]` **The U2 seam review earned its place, and how** — inline.place indexes by subscript — the miss a cold panel cannot see → `persistent-memory.d/2026-09-21-u2-seam-review-earned-it.md`
-  SECOND subscript in `inline.place` the seam review undercounted, and a
+- `[2026-09-21]` **Marks are one `.marks.json` per booth** — operator decision with two rejected alternatives; read before restructuring → `persistent-memory.d/2026-09-21-marks-storage-decision.md`
-  regression in my own legacy importer that a retargeted test caught — a
+- `[2026-09-21]` **U7's section premise is half wrong** — every booth that needs navigation is FLAT — read before starting U7 → `persistent-memory.d/2026-09-21-u7-section-premise-half-wrong.md`
-  malformed sidecar that renders `⚠ broken` today would have silently vanished
+- `[2026-09-21]` **`sindra-finalists` is U2's flag motivation, caught live** — evidence, not argument → `persistent-memory.d/2026-09-21-sindra-finalists-is-the-motivation.md`
-  on migration.
+- `[2026-09-21]` **The information architecture and the v1 gate landed** — the single defect the seven units decompose → `persistent-memory.d/2026-09-21-ia-and-v1-gate-landed.md`
-- `[2026-09-21]` **Marks are stored as one `.marks.json` per booth**, atomic
+- `[2026-09-21]` **The `.forever` diagnosis is a falsifiable prediction** — U4's success criterion — re-count a fortnight AFTER U4 lands → `persistent-memory.d/2026-09-21-forever-diagnosis-is-a-prediction.md`
-  temp-file + `os.replace`, `fcntl` lock on the read-modify-write — operator
+- `[2026-09-21]` **Extracted from `eshpfi` into its own repo** — test_booth.py is the regression net the v1 rewrite is checked against → `persistent-memory.d/2026-09-21-extracted-from-eshpfi.md`
-  decision, this session. Two alternatives were weighed and lost: a sidecar
-  per item (`<rel>.marks.json`) and extending the existing `<stem>.ask.json`
-  shape. Rationale, and the reason it is not `links.md`-shaped: **(a)** U4
-  makes *"does this booth owe an answer?"* a hot question — the sweep asks it
-  per booth per tick and the index asks it per card per page load, so per-item
-  sidecars turn it into a full walk of all 25 booths, one of which holds 270
-  files; **(b)** `links.md` is an `O_APPEND` content-hash log because **17
-  agent handles write it concurrently**, whereas marks have exactly one writer
-  (the operator, in one browser) and many readers — a different problem that
-  must not inherit the append-log design; **(c)** `.blurred` / `.pins` /
-  `.forever` already establish the per-booth dotfile as the house shape for
-  operator state, and `booth_items()`'s dotfile skip means it costs nothing in
-  counts, galleries or zips. Accepted cost: a corrupt `.marks.json` loses that
-  booth's marks rather than one item's. Implementation deferred to U2 —
-  tracked at `ROADMAP.md` U2 and by this entry.
-- `[2026-09-21]` **U7's section premise is half wrong, and it is the half that
-  matters** — found by re-measuring `~/booth-data` rather than trusting the IA
-  doc. The IA says sections come from subfolders that already exist on disk;
-  true, but **every booth that actually needs navigation is flat**:
-  `pancake-v3-full` (270 items, 0 subfolders), `pancake-v4-full` (270, 0),
-  `sindra20-engines` (98 items + 99 caption sidecars, 0), `sindra-finalists`
-  (86 + 87, 0). Subfolders exist on exactly two booths — `pewpew-ui-brief` (7,
-  nested to `_ds/powerpellet-design-system-<uuid>/preview`) and `dfa-concepts`
-  (1) — and **both are reports**, the job where grid navigation matters least.
-  So sections stay worth shipping and `Item.section` stays right, but they are
-  **not** "most of the navigation fix": the rail, the filters and grid keyboard
-  are all of it. Worth noting for whoever writes U7: `sindra20-engines` encodes
-  its structure in the **filename prefix** (`b2-s1-<subject>-<seed>`), which is
-  where a grouping heuristic would actually pay. The IA doc's claim about what
-  sections buy needs a line struck — not yet edited.
-- `[2026-09-21]` **`sindra-finalists` is U2's `flag` motivation caught in the
-  act** — 86 items, every one captioned, and the booth's entire name is "the
-  ones the operator picked." That loop currently runs through chat, which is
-  the defect `flag` closes. Evidence, not argument.
-- `[2026-09-21]` **The information architecture and the v1 gate landed**
-  (`726822b`): `docs/design/information-architecture.md` names the single
-  defect — *one lifetime (24h from last touch) and one shape (a folder),
-  serving five jobs with different lifetimes and different shapes* — and
-  `ROADMAP.md` gates v1 on seven units, each closing a **measured** defect
-  rather than a wish. Both were written after a measurement pass over the live
-  service, and the measurements are the load-bearing part.
-- `[2026-09-21]` **The `.forever` diagnosis is a stated, falsifiable
-  prediction.** U4 (derived lifetime) predicts the kept-rate falls to the
-  genuinely-durable booths. Re-measured today: **14 of 25 booths kept (56%)**,
-  against the 54% the IA doc recorded. **Re-count a fortnight after U4 lands.**
-  If it does not move, the diagnosis was wrong and the boolean was doing
-  something else. Tracked in the IA doc's Booth section and by this entry.
-- `[2026-09-21]` **Extracted from `eshpfi` into its own repo.** The accreted
-  service came over whole, tests included, so `tests/test_booth.py` (1581 lines)
-  is the regression net the v1 rewrite is checked against.
 
 ## Tried and abandoned
 
-- `[2026-09-21]` **Letting Jinja hot-reload templates while the repo is the
+- `[2026-09-21]` **Tagging a release while a review gate was in flight** — cost a same-hour v0.2.1 and a correction to 15 handles → `persistent-memory.d/2026-09-21-tagging-with-a-gate-in-flight.md`
-  deployment root** — the cause of a live outage the same day U2 landed, and the
+- `[2026-09-21]` **Letting the write path share the read path's leniency** — a tolerant reader and a tolerant writer are not the same decision → `persistent-memory.d/2026-09-21-tolerant-writer-over-tolerant-reader.md`
-  sharpest foot-gun in the repo. `booth.service` sets `WorkingDirectory` to this
+- `[2026-09-21]` **Letting Jinja hot-reload templates in the deployment root** — caused a live outage: 19 of 25 booths at 500. Why auto_reload=False → `persistent-memory.d/2026-09-21-jinja-hot-reload-outage.md`
-  repo, so the running service imports these files with no build step and no
+- `[2026-09-21]` **Five mechanisms to get one question beside one artifact** — the accretion signature this whole v1 rewrite is undoing → `persistent-memory.d/2026-09-21-five-mechanisms-one-job.md`
-  staging copy. Python is read once at process start; Jinja's `FileSystemLoader`
+- `[2026-09-21]` **Regex-injecting chrome into arbitrary author HTML** — the defect U3 exists to close → `persistent-memory.d/2026-09-21-regex-injecting-chrome.md`
-  re-reads a template **on every render**. Editing `booth.html` therefore
+- `[2026-09-21]` **A boolean escape hatch as the lifetime mechanism** — why `.forever` is a symptom; the defect U4 exists to close → `persistent-memory.d/2026-09-21-boolean-escape-hatch-as-lifetime.md`
-  deployed it instantly against Python from 22:03 that knew nothing about
+- `[2026-09-21]` **Letting the link board absorb the announce job** — 69% rot; U5 gave the job a home, which is what unblocks U6 → `persistent-memory.d/2026-09-21-link-board-absorbing-announce.md`
-  `item_marks`, and **19 of 25 live booths returned 500** with
-  `UndefinedError: 'item_marks' is undefined`. Neither the old code nor the new
-  code was broken — the service was running both at once.
-  **The lesson that generalises:** a skew between a process and the disk under it
-  is invisible to the test suite by construction, so no amount of green tests
-  would have caught it; the operator found it. Fixed at the source rather than
-  with a reminder — the `Environment` is hand-built with `auto_reload=False`, so
-  there is now ONE staleness rule (nothing takes effect until you restart) and
-  the running process is always a coherent snapshot of one commit. Asserted by
-  `test_templates_do_not_hot_reload_from_disk`. Watch the second-order risk the
-  fix introduces: a hand-built `Environment` does not inherit `autoescape` from
-  the `Jinja2Templates` constructor, and booth names, item names and mark text
-  are all agent-authored strings landing in HTML.
-
-- `[2026-09-21]` **Five separate mechanisms to get one question next to one
-  artifact** — `.forever`, the link board, `inline.py`'s placeholder DSL,
-  `wrap_verbatim_html`'s six regexes, and the floating amber asks chip plus
-  `/b/<n>/asks`. Every one is a *correct local fix* to the same global
-  mismatch, which is exactly why they accumulated without anyone making a bad
-  call. **The foot-gun is the sixth one:** the next "just add a small thing for
-  this case" reads as reasonable and is the pattern. The git log carries the
-  signature — every feature ships, then takes 2–5 patches for cases the single
-  shape did not anticipate. Check the ROADMAP gate before adding a mechanism.
-- `[2026-09-21]` **Regex-injecting chrome into arbitrary author HTML**
-  (`wrap_verbatim_html` + `_HEAD_CLOSE_RE`, `_HTML_OPEN_RE`, `_DOCTYPE_RE`,
-  `_BODY_CLOSE_RE`, `_HTML_CLOSE_RE`, `_ICON_RE`, and the doctype/charset
-  ordering constraints they thread). It works today and is **still live** —
-  but it is the single most fragile thing in the service and it is load-bearing
-  for the operator's most important workflow. Slated for deletion at U3 in
-  favour of a declared seam (`/_booth/embed.js`, mounted through a real DOM
-  API), which costs an author one line and removes the whole class. Do not
-  extend the regex set in the meantime; if a verbatim page breaks, that is an
-  argument for U3, not for a seventh pattern.
-- `[2026-09-21]` **A boolean escape hatch as the lifetime mechanism.**
-  `.forever` was added because a 24h TTL genuinely did not fit some booths —
-  and then 56% of live booths ended up on it, which means it is not "ephemeral
-  with an exception", it is two lifetimes wearing one lifetime's clothes, with
-  the operator doing the sorting by hand. Replaced at U4 by lifetime derived
-  from state (an open mark pins; viewing is activity; `keep` survives as an
-  explicit reasoned pin rather than the only way to say "not yet").
-- `[2026-09-21]` **Letting the link board absorb the announce job.** `booth
-  link` is an `O_APPEND` write with no identity and no stated rule, so
-  re-announcing a bench appends a row instead of updating one, and a booth URL
-  rots the moment its booth is swept — **145 of 211 rows (69%) pointed at
-  nothing**, and 22 were the same target re-posted (talk 5×, peedlar 4×). The
-  rot is **structural, not drift**. The lesson that cost the most: enforcing
-  the link rule without first giving the announce job a home (`.booth.json`
-  provenance on the index, U5) just makes it homeless.

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "booth"
-version = "0.2.1"
+version = "1.0.0b1"
 description = "The Booth — a dead-simple standing web server that scans a data dir of drop-folders and renders each as an ephemeral media 'booth' (image/webm/audio auto-gallery, or a folder's own index.html verbatim). Also accepts browser/curl uploads for pickup under a human-readable id. 24h TTL, then the folder is wiped. Fleet tool for CC sessions to surface A/B and smoke results to the operator."
 requires-python = ">=3.11"
 dependencies = [
@@ -15,6 +15,15 @@ dependencies = [
 test = [
     "pytest>=8.0",
     "httpx>=0.27",  # fastapi TestClient
+    # U3's embed seam moves placement into the browser, where no string
+    # assertion can see it. Browsers are NOT downloaded per project: they live
+    # box-wide in /opt/ms-playwright with PLAYWRIGHT_BROWSERS_PATH wired
+    # globally. THE UPPER BOUND IS THE POINT -- each playwright release pins a
+    # Chromium revision, and one that wants a revision the shared store does
+    # not have dies with an opaque "Executable doesn't exist" rather than a
+    # missing-dependency error. 1.60-1.62 map to chromium 1223/1228/1234, all
+    # present. Raise the bound only after the store has the newer revision.
+    "playwright>=1.60,<1.63",
 ]
 
 [build-system]

scripts/booth

@@ -3,18 +3,46 @@
 # folder under $BOOTH_DATA_DIR; this is sugar over mkdir/cp so you get the URL
 # back.
 #
-#   booth new    <name>                 make an empty booth, print its URL
+#   booth new    <name> [--why W] [--title T]
-#   booth add    <name> <file>...       copy files into a booth (creates it), print URL
+#                                       make an empty booth, print its URL
+#   booth add    <name> <file>... [--why W] [--title T]
+#                                       copy files into a booth (creates it), print URL
 #   booth url    <name>                 print a booth's URL
 #   booth ls                            list booths (kept ones marked ★)
 #   booth rm     <name>                 wipe a booth now (TTL would eventually anyway)
 #
 #   booth keep   <name>                 exempt a booth from the 24h sweep, forever
+#                                       (NOT for "waiting on an answer" — an open
+#                                        pick holds its own booth, see below)
 #   booth unkeep <name>                 hand it back to the sweeper
 #   booth link   <url> [description]    append a link to the standing link board
+#                                       REFUSES a booth URL — a booth announces
+#                                       itself now; use `booth new --why`
 #   booth links                         list the board, numbered, with entry ids
 #   booth unlink <id|index>             remove ONE link from the board
 #
+#   booth bench add <url> <name>        register or UPDATE a bench (upsert)
+#   booth bench ls                      list benches, live -> promoted -> retired
+#   booth bench state <id|url> <state>  live | promoted | retired
+#   booth bench rm <id|url>             remove one
+#   booth bench import                  classify the board's rows; writes NOTHING
+#   booth bench import --apply <id>...  register ONLY the ids you name. A bare
+#                                       --apply is REFUSED: a machine cannot tell
+#                                       a bench from a bookmark by its URL, and
+#                                       ~14 of 35 live candidates are bookmarks.
+#                                       links.md is never edited by either form.
+#
+# THREE SURFACES, THREE JOBS. Telling them apart is the whole of U6:
+#   a BOOTH is a review surface you post work to. It announces itself and is
+#     swept 24h after its last activity. `booth new` / `booth add`.
+#   a BENCH is a running thing — jackdaw's bench, talk's bench, the things that
+#     get promoted to Homepage. Durable, and identified BY ITS URL, so posting
+#     it again updates the row instead of adding a fifth. `booth bench add`.
+#   a LINK is a reference bookmark — a repo, a model card, a doc page. The
+#     standing board, unchanged and NOT deprecated. `booth link`.
+# The board carried all three because only one of them had a surface: 178 of its
+# 221 rows were booth URLs and 156 of those pointed at booths already swept.
+#
 #   booth ask    <name> <id> <prompt> <option>... [--no-notes]
 #                                       pose a multiple-choice question in a booth
 #   booth marks  <name> [--wait [SECS]] print every mark in a booth as JSON;
@@ -22,6 +50,17 @@
 #   booth answer <name> <id> [--wait [SECS]]
 #                                       print ONE pick's answer (exit 1 if unanswered);
 #                                       --wait polls until it lands (default 3600 s)
+#
+# EXIT CODES for the two reading verbs. A read that FAILED gets its own code so
+# a caller can tell "not yet" from "the file is damaged" — conflating them is
+# how a broken `.marks.json` used to look like an unanswered question and wait
+# out the full hour.
+#   marks   0 read ok · 1 --wait timed out with picks open · 3 unreadable
+#   answer  0 answered · 1 unanswered · 2 no such pick · 3 unreadable ·
+#           4 the pick hydrated broken and can never be answered
+#
+# `answer` and `marks` use the SAME openness predicate. A partially-answered
+# pick is still open to both; a broken one is closed to both.
 #   booth marks-import <name>           import legacy *.ask.json into .marks.json
 #   booth asks   <name>                 alias for `marks` (deprecated)
 #
@@ -46,12 +85,28 @@
 # access, so they poll the HTTP mirror instead:
 #   http://10.100.10.50:8090/b/<name>/marks.json
 #
-# THE 24h RULE AND ITS ONE EXCEPTION. Every booth is wiped 24h after its last
+# THE 24h RULE AND ITS THREE STATES. Every booth is wiped 24h after its last
 # activity — that is the contract, and it is why nobody has to clean up after
-# themselves. `keep` drops a `.forever` sentinel that exempts one booth from the
+# themselves. Two things exempt a booth, and only the first is a button:
-# sweep and moves it into its own lane at the top of the index. Use it for
+#
-# durable operator-facing boards, not for run output. `unkeep` is just `rm` of
+#   KEPT  `keep` drops a `.forever` sentinel that exempts one booth from the
-# the sentinel, so putting a board back under the sweeper costs nothing.
+#         sweep and moves it into its own lane at the top of the index. Use it
+#         for durable operator-facing boards, not for run output. `unkeep` is
+#         just `rm` of the sentinel, so putting a board back costs nothing.
+#   HELD  a booth with an UNANSWERED pick is never swept, automatically, for as
+#         long as the question is open. You do not press anything: `booth ask`
+#         is what holds it, and the operator answering is what releases it. A
+#         partially-answered pick still counts as open, so a review in flight
+#         cannot be swept out from under him.
+#
+# So: DO NOT `keep` a booth just because you are waiting on an answer. That was
+# the old workaround, it is what made 70% of live booths "durable", and it is
+# no longer needed. `keep` means durable. The question holds its own booth.
+#
+# VIEWING IS ACTIVITY TOO. The operator opening a booth page resets its clock —
+# if he is still looking at it, it is still alive. Your polling does NOT: `booth
+# marks --wait` and the `marks.json` endpoint are machine reads and deliberately
+# do not count, so a session cannot hold its own booth open by waiting on it.
 #
 # DELETING A KEPT BOARD: `booth rm <name>` works on kept boards too and deletes
 # NOW — it announces that the board was kept, so wiping something durable is
@@ -59,15 +114,29 @@
 # card drops the sentinel, the card moves to the ephemeral lane, and the × wipes
 # it from there.
 #
-# DO NOT "unkeep and let it expire". Removing the sentinel BUMPS the booth
+# DO NOT "unkeep and let it expire". RELEASING A BOARD IS ACTIVITY — you just
-# directory's mtime, and a booth's age is the newest mtime in its tree — so a
+# touched it — so a released board's clock resets and it survives another full
-# released board's clock RESETS and it survives another full 24h. Unkeep-and-wait
+# 24h. Unkeep-and-wait is a delay, not a delete. Use `rm` (or the UI ×) when you
-# is a delay, not a delete. Use `rm` (or the UI ×) when you mean now.
+# mean now. (This was true before U4 as an accident of directory mtime; it is
+# now the stated rule, which is why it no longer needs a warning shaped like a
+# surprise.)
 #
 # `link` is the reason the exception exists: agent sessions hand the operator
 # URLs that then drown in terminal scrollback. They go on a standing kept board
 # instead, with provenance, so they outlive the session that produced them.
 #
+# ANNOUNCE YOUR BOOTH. `--why` is one line saying what the operator is looking
+# at and why he should care; it lands on the index card and on the booth page
+# beside your handle, taken from $ALTHING_HANDLE. It is optional and nothing
+# breaks without it — but a booth that cannot say what it is has no way to ask
+# for attention except by posting its URL somewhere, which is exactly how the
+# link board came to be 69% dead rows. The booth is the place to say it.
+#
+#   booth add r18-ab out/*.png --why "pick the denoiser, left column is v3"
+#
+# Re-announcing (a second `new` or `add` on the same booth) updates the why and
+# KEEPS the original creation stamp: the booth appeared once.
+#
 # On a host that is NOT nh3-dev, rsync into the data dir instead, e.g.:
 #   rsync -a ./out/  nh3-dev:booth-data/my-run/
 set -euo pipefail
@@ -78,23 +147,121 @@ KEEP=".forever"                                    # must match KEEP_MARKER in b
 BLUR=".blurred"   # one booth-relative item path per line; see `blur` below
 LINKS_BOARD="${BOOTH_LINKS_BOARD:-links}"
 
+# `--why` / `--title` for `new` and `add`. Pulled out of "$@" wherever they
+# appear, so `booth add b *.png --why "..."` and `booth add b --why "..." *.png`
+# both work — a glob is usually last and a flag usually after it, but nothing
+# enforces that and a session should not have to care.
+# OMITTED IS NOT EMPTY. `booth new x --why "..."` then `booth add x out/*.png`
+# is the ordinary sequence, and while an omitted flag meant "" the second
+# command silently erased the sentence the first one existed to record. So the
+# shell tracks WHETHER the flag was given, and only passes it on when it was —
+# an explicit `--why ""` still clears, which is a different intention.
+WHY=""; TITLE=""; WHY_SET=0; TITLE_SET=0; ARGS=()
+strip_announce_flags() {
+  ARGS=(); WHY_SET=0; TITLE_SET=0
+  while [ $# -gt 0 ]; do
+    case "$1" in
+      --why)     [ $# -ge 2 ] || usage; WHY="$2";   WHY_SET=1;   shift 2 ;;
+      --title)   [ $# -ge 2 ] || usage; TITLE="$2"; TITLE_SET=1; shift 2 ;;
+      --why=*)   WHY="${1#--why=}";     WHY_SET=1;   shift ;;
+      --title=*) TITLE="${1#--title=}"; TITLE_SET=1; shift ;;
+      *) ARGS+=("$1"); shift ;;
+    esac
+  done
+}
+
+# Announce a booth. Goes through booth/manifest.py rather than printf-ing JSON
+# from the shell, because a why containing a quote, a backslash or a newline is
+# not an edge case — it is a sentence somebody wrote.
+# announce <dir> <handle> [title] [why] — the trailing two are passed as
+# environment variables that are UNSET when the flag was not given, because
+# that is the only way the shell can say "leave it alone" rather than "".
+announce() {
+  local -a envs
+  envs=( "BOOTH_SRC=$(cd "$(dirname -- "$(readlink -f -- "$0")")/.." && pwd)"
+         "BOOTH_ANN_DIR=$1" "BOOTH_ANN_HANDLE=$2" )
+  [ "${TITLE_SET:-0}" = 1 ] && envs+=( "BOOTH_ANN_TITLE=${3:-}" )
+  [ "${WHY_SET:-0}" = 1 ]   && envs+=( "BOOTH_ANN_WHY=${4:-}" )
+  env "${envs[@]}" python3 -c '
+import os, pathlib, sys
+sys.path.insert(0, os.environ["BOOTH_SRC"])
+try:
+    from booth.manifest import write_manifest
+    kw = {}
+    # Absent means the flag was omitted; present-and-empty means it was given
+    # as "" and the poster meant to take the line back.
+    if "BOOTH_ANN_TITLE" in os.environ: kw["title"] = os.environ["BOOTH_ANN_TITLE"]
+    if "BOOTH_ANN_WHY" in os.environ:   kw["why"] = os.environ["BOOTH_ANN_WHY"]
+    write_manifest(pathlib.Path(os.environ["BOOTH_ANN_DIR"]),
+                   os.environ["BOOTH_ANN_HANDLE"], **kw)
+except Exception as exc:
+    # A booth that could not announce itself is still a booth. Say so on stderr
+    # and carry on: failing `booth add` over its metadata would lose the files
+    # the session just copied, which is a far worse trade.
+    print(f"booth: could not write the announcement: {exc}", file=sys.stderr)
+'
+}
+
+# Who is posting. The same chain `link` uses for its rows, so provenance means
+# the same thing on the board and on the card.
+whoami_handle() {
+  echo "${ALTHING_HANDLE:-${BOOTH_SOURCE:-$(hostname -s 2>/dev/null || echo unknown)}}"
+}
+
+# Where booth/*.py lives, for the `python3 -c` calls below. The CLI runs under
+# the SYSTEM python3 with no venv, which is why every module it imports is
+# stdlib-only (CLAUDE.md invariant 1) and why no AST extractor can see these
+# imports — `tests/test_cli.py` runs the real script, and is the only thing that
+# catches a third-party import before a fleet host does.
+booth_src() {
+  (cd "$(dirname -- "$(readlink -f -- "$0")")/.." && pwd)
+}
+
+# The booth NAME a URL points at, or empty. ONE PREDICATE — this shells out to
+# booth.links.booth_target rather than pattern-matching `:8090/b/` here, because
+# the board's dead-row marker and `bench import` use that same function and a
+# second implementation in the shell would classify the host-agnostic and
+# percent-encoded cases differently (INV-2).
+# Prints `B:<name>` for a booth URL and `N` for anything else.
+#
+# A SENTINEL, NOT AN EMPTY STRING. Command substitution strips trailing
+# newlines, so a predicate that answers with the bare name cannot distinguish
+# "not a booth" from "a booth whose name bash just erased" — and the guard
+# then fails OPEN on that edge, which is the one direction a guard must never
+# fail. The prefix makes the answer unambiguous whatever the name contains.
+booth_target_of() {
+  BOOTH_SRC="$(booth_src)" BOOTH_Q="$1" python3 -c '
+import os, sys
+sys.path.insert(0, os.environ["BOOTH_SRC"])
+from booth.links import booth_target          # stdlib only — no venv needed
+name = booth_target(os.environ["BOOTH_Q"])
+sys.stdout.write("N" if name is None else "B:" + name)
+'
+}
+
 usage() {
-  echo "usage: booth {new <name>|add <name> <file>...|url <name>|ls|rm <name>|keep <name>|unkeep <name>|blur <name> <file>...|unblur <name> <file>...|link <url> [description]|links|unlink <id|index>|ask <name> <id> <prompt> <option>... [--no-notes]|marks <name> [--wait [SECS]]|answer <name> <id> [--wait [SECS]]|marks-import <name>}" >&2
+  echo "usage: booth {new <name> [--why W] [--title T]|add <name> <file>... [--why W] [--title T]|url <name>|ls|rm <name>|keep <name>|unkeep <name>|blur <name> <file>...|unblur <name> <file>...|link <url> [description]|links|unlink <id|index>|ask <name> <id> <prompt> <option>... [--no-notes]|marks <name> [--wait [SECS]]|asks <name> (deprecated alias for marks)|answer <name> <id> [--wait [SECS]]|marks-import <name>|bench add <url> <name>|bench ls|bench state <id|url> <live|promoted|retired>|bench rm <id|url>|bench import [--apply <id>...]}" >&2
   exit 2
 }
 
 cmd="${1:-}"; shift || true
 case "$cmd" in
   new)
+    strip_announce_flags "$@"
+    set -- ${ARGS+"${ARGS[@]}"}
     [ $# -ge 1 ] || usage
     mkdir -p -- "$DATA/$1"
+    announce "$DATA/$1" "$(whoami_handle)" "$TITLE" "$WHY"
     echo "$URL/b/$1/"
     ;;
   add)
+    strip_announce_flags "$@"
+    set -- ${ARGS+"${ARGS[@]}"}
     [ $# -ge 2 ] || usage
     name="$1"; shift
     mkdir -p -- "$DATA/$name"
     cp -- "$@" "$DATA/$name/"
+    announce "$DATA/$name" "$(whoami_handle)" "$TITLE" "$WHY"
     echo "$URL/b/$name/"
     ;;
   url)
@@ -167,9 +334,65 @@ case "$cmd" in
     [ $# -ge 1 ] || usage
     link_url="$1"; shift
     desc="${*:-}"
+    # THE REFUSAL COMES FIRST, BEFORE ANY WRITE (INV-3). A booth announces
+    # itself now (U5), so a booth URL on the board is a row that rots the
+    # moment the booth is swept — 156 of the board's 221 rows are exactly
+    # that. Refusing AFTER the mkdir/announce below would leave a new booth
+    # behind as the side effect of a call that failed.
+    # `|| pred_rc=$?` so a BROKEN PREDICATE is handled here rather than aborting
+    # the script under `set -e` with a raw Python traceback and nothing else.
+    # The direction is FAIL-CLOSED and stays that way: if we cannot tell whether
+    # this is a booth, we do not append. A guard that fails open is not a guard,
+    # and the cost of being wrong in the other direction is one message telling
+    # the poster exactly what broke.
+    pred_rc=0
+    refused_name="$(booth_target_of "$link_url" 2>/dev/null)" || pred_rc=$?
+    if [ "$pred_rc" -ne 0 ]; then
+      {
+        echo "booth link: could not check whether that URL is a booth, so nothing was posted."
+        echo "  the check runs booth/links.py under the system python3 with no venv."
+        echo "  re-run from a checkout where \`python3 -c 'import booth.links'\` works,"
+        echo "  or post it from a host that has one."
+      } >&2
+      exit 3
+    fi
+    case "$refused_name" in
+      N) refused_name="" ;;
+      B:*) refused_name="${refused_name#B:}" ;;
+      *)
+        echo "booth link: the booth check answered something unrecognised; nothing was posted." >&2
+        exit 3 ;;
+    esac
+    # CREDENTIALS DO NOT GO ON THE BOARD, through any door. `normalize_bench_url`
+    # refuses userinfo for a bench; `booth link` is the door this unit did not
+    # touch, and the board renders on an unauthenticated LAN surface. A small,
+    # deliberate widening of the unit -- named rather than smuggled.
+    case "$link_url" in
+      *://*@*)
+        {
+          echo "booth link: that URL carries credentials (user:pass@host) and the board"
+          echo "  is readable by anyone who can reach this service. Nothing was posted."
+          echo "  strip the credentials and post it again."
+        } >&2
+        exit 2 ;;
+    esac
+    if [ -n "$refused_name" ]; then
+      {
+        echo "booth link: that is a booth, and a booth announces itself now."
+        echo "  booth new $refused_name --why \"${desc:-what the operator is looking at}\""
+        echo "  (or --why on \`booth add\`; re-announcing keeps the original stamp)"
+        echo "  the index at $URL/ is the feed."
+      } >&2
+      exit 2
+    fi
     board="$DATA/$LINKS_BOARD"
     mkdir -p -- "$board"
     : > "$board/$KEEP"                             # the board is durable by definition
+    # The board announces itself as the SERVICE's, not as any one agent's:
+    # seventeen handles post to it, so no handle owns it. Idempotent — a second
+    # link keeps the original creation stamp.
+    TITLE_SET=1 WHY_SET=1 announce "$board" "booth" "$LINKS_BOARD" \
+      "the standing link board — every agent session posts here"
     # Provenance, because a bare URL is unreadable three days later: who posted
     # it, from where, and when.
     who="${ALTHING_HANDLE:-${BOOTH_SOURCE:-$(hostname -s 2>/dev/null || echo unknown)}}"
@@ -182,9 +405,20 @@ case "$cmd" in
     # shared lock this line could land inside that window and be rewritten
     # away by the prune.
     touch -- "$board/.links.lock"
-    flock "$board/.links.lock" \
+    # THE REDIRECTION OPENS INSIDE THE LOCK, which is why this is `sh -c` and
-    printf -- '- [%s](%s) <sub>· %s · %s</sub>\n' \
+    # not a bare printf. `flock LOCK printf ... >> board` reads as locked and is
-      "${desc:-$link_url}" "$link_url" "$who" "$when" >> "$board/links.md"
+    # not: the SHELL opens the append fd while parsing, before flock acquires
+    # anything. If a concurrent `unlink` rewrites the board in that window, the
+    # rewrite lands on a NEW inode via os.replace and this fd still points at
+    # the old, unlinked one — so the append succeeds, reports success, and the
+    # row is gone. Found by a cold bug-hunt arm; pre-existing, not U6's, but it
+    # is a silent data loss in the file this unit spends its time in.
+    BK_DESC="${desc:-$link_url}" BK_URL="$link_url" BK_WHO="$who" BK_WHEN="$when" \
+    BK_BOARD="$board/links.md" \
+    flock "$board/.links.lock" sh -c '
+      printf -- "- [%s](%s) <sub>· %s · %s</sub>\n" \
+        "$BK_DESC" "$BK_URL" "$BK_WHO" "$BK_WHEN" >> "$BK_BOARD"
+    ' 
     echo "$URL/b/$LINKS_BOARD/"
     ;;
   links)
@@ -240,6 +474,149 @@ if removed is None:
 print("removed: %s  %s" % (removed["desc"], removed["url"]))
 ' "$board" "$target"
     ;;
+  bench)
+    # SEAM REVIEW SR-6: the first two-word verb in this script. A nested case,
+    # and a bare `bench` names the bench verbs rather than falling through to
+    # the generic usage, which would hide which of the two words was wrong.
+    sub="${1:-}"; shift || true
+    case "$sub" in
+      add|ls|state|rm|import) ;;
+      *)
+        echo "usage: booth bench {add <url> <name>|ls|state <id|url> <live|promoted|retired>|rm <id|url>|import [--apply <id>...]}" >&2
+        exit 2 ;;
+    esac
+    BOOTH_SRC="$(booth_src)" BOOTH_DATA="$DATA" BOOTH_SUB="$sub" \
+    BOOTH_WHO="$(whoami_handle)" BOOTH_BOARD="$LINKS_BOARD" \
+    python3 -c '
+import os, pathlib, sys
+sys.path.insert(0, os.environ["BOOTH_SRC"])
+# stdlib only — no venv needed. benches.py imports no sibling either (INV-9).
+from booth.benches import (normalize_bench_url, order_benches, read_benches,
+                           remove_bench, set_bench_state, upsert_bench)
+from booth.links import booth_target, parse_link_entries
+
+root = pathlib.Path(os.environ["BOOTH_DATA"])
+sub, who = os.environ["BOOTH_SUB"], os.environ["BOOTH_WHO"]
+argv = sys.argv[1:]
+
+def die(msg, code=2):
+    print("booth bench: %s" % msg, file=sys.stderr)
+    raise SystemExit(code)
+
+def row(b):
+    # ONE LINE PER BENCH, in the rendered order — state first, so a retired
+    # bench sinks, then name, then id as a total tie-break (INV-4).
+    # THE ID IS PRINTED WHOLE AND UNTRUNCATED, because it is the locator
+    # `bench state` and `bench rm` take: a truncated one is not an id, it is a
+    # string that looks like one and silently addresses nothing. The name and
+    # the added date are the truncatable columns.
+    return "%-9s  %-10s  %-16s  %-24s  %s" % (
+        b.state, b.added[:10], b.owner[:16], b.name[:24], b.id)
+
+if sub == "add":
+    if len(argv) < 2: die("bench add <url> <name>")
+    try:
+        bench, created = upsert_bench(root, argv[0], " ".join(argv[1:]), who)
+    except ValueError as exc:
+        die(exc)
+    print("%s: %s" % ("registered" if created else "updated", bench.id))
+elif sub == "ls":
+    benches, err = read_benches(root)
+    if err:
+        die("the registry could not be read: %s" % err, 3)
+    if not benches:
+        print("no benches registered yet")
+    else:
+        print("%-9s  %-10s  %-16s  %-24s  %s"
+              % ("STATE", "ADDED", "OWNER", "NAME", "ID (pass to state|rm)"))
+    for b in benches:
+        print(row(b))
+elif sub in ("state", "rm"):
+    if not argv: die("bench %s <id|url>%s" % (sub, " <state>" if sub == "state" else ""))
+    try:
+        bench_id = normalize_bench_url(argv[0])
+    except ValueError as exc:
+        die(exc)
+    if sub == "rm":
+        gone = remove_bench(root, bench_id)
+        if gone is None: die("no such bench: %s" % bench_id, 1)
+        print("removed: %s" % gone.url)
+    else:
+        if len(argv) < 2: die("bench state <id|url> <live|promoted|retired>")
+        try:
+            moved = set_bench_state(root, bench_id, argv[1])
+        except ValueError as exc:
+            die(exc)
+        if moved is None: die("no such bench: %s" % bench_id, 1)
+        print("%s is now %s" % (moved.url, moved.state))
+elif sub == "import":
+    apply = "--apply" in argv
+    picked = [a for a in argv if a != "--apply"]
+    board = root / os.environ["BOOTH_BOARD"] / "links.md"
+    if not board.is_file(): die("no link board at %s" % board, 1)
+    skipped, candidates, refused = [], [], []
+    for e in parse_link_entries(board.read_text()):
+        name = booth_target(e["url"])
+        if name is not None:
+            skipped.append((e, name)); continue
+        try:
+            candidates.append((normalize_bench_url(e["url"]), e))
+        except ValueError as exc:
+            refused.append((e, str(exc)))
+    print("SKIPPED — booth rows; a booth announces itself now (%d):" % len(skipped))
+    for e, name in skipped:
+        print("  %-30s %s" % (name, e["url"]))
+    print()
+    print("CANDIDATES — would be registered (%d rows, %d distinct):"
+          % (len(candidates), len({i for i, _ in candidates})))
+    for i, e in candidates:
+        # THE NORMALIZED ID BESIDE THE RAW URL, which is the whole point of the
+        # proposal: five rows of `talk` collapsing to one is only visible if you
+        # can see which five raw URLs produced the one id. The description is
+        # the thing to drop here, not the URL.
+        print("  %-52s %s" % (i, e["url"]))
+        if e["desc"]:
+            print("  %-52s   %s" % ("", e["desc"][:70]))
+    print()
+    print("REFUSED — normalization said no (%d):" % len(refused))
+    for e, why in refused:
+        print("  %-52s %s" % (e["url"], why))
+    if not apply:
+        print()
+        print("nothing was written.")
+        print("  booth bench import --apply <id>...   register ONLY the ids you name")
+        print()
+        print("A MACHINE CANNOT TELL A BENCH FROM A BOOKMARK BY ITS URL. On the live")
+        print("board roughly 14 of 35 candidates are repos, model cards and docs, for")
+        print("which the board is the right and only home. So `--apply` takes the ids")
+        print("YOU pick from the list above; it will not register the whole set.")
+        raise SystemExit(0)
+    # SELECTION IS MANDATORY. A bare `--apply` would do exactly the thing this
+    # rationale of this very unit says is impossible -- decide bench-vs-bookmark
+    # URL -- and it would do it silently, to ~14 rows that belong on the board.
+    # The dry-run prints the ids; the operator names the ones that are benches.
+    if not picked:
+        print()
+        print("booth bench import --apply needs the ids to register.", file=sys.stderr)
+        print("  nothing was written. copy the ids you want from the list above:",
+              file=sys.stderr)
+        print("  booth bench import --apply <id> [<id>...]", file=sys.stderr)
+        raise SystemExit(2)
+    by_id = {i: e for i, e in candidates}
+    unknown = [i for i in picked if i not in by_id]
+    if unknown:
+        print()
+        for i in unknown:
+            print("not a candidate id: %s" % i, file=sys.stderr)
+        print("nothing was written.", file=sys.stderr)
+        raise SystemExit(2)
+    for i in picked:
+        e = by_id[i]
+        upsert_bench(root, e["url"], e["desc"], e["who"] or who)
+    print()
+    print("applied: %d bench(es) registered. links.md was NOT modified." % len(set(picked)))
+' "$@"
+    ;;
   ask)
     # booth ask <name> <id> <prompt> <opt>... [--no-notes]
     [ $# -ge 5 ] || usage
@@ -270,6 +647,15 @@ except MarksCorrupt as exc:
     ;;
   marks|asks)
     # booth marks <name> [--wait [SECS]]     (`asks` is the deprecated alias)
+    #
+    # EXIT CODES. 0 = the read succeeded and the document is on stdout; 1 =
+    # --wait gave up with picks still open (the document is still printed); 3 =
+    # the marks could not be read at all. A reader that CRASHED must never look
+    # like an answer — the old shape printed a traceback and exited 0, so a
+    # caller piping to `jq` saw success and got nothing.
+    #
+    # Whether anything is still open is in the payload's `open` list. The read
+    # verb does not encode it in its status: a successful read is a success.
     [ $# -ge 1 ] || usage
     name="$1"; shift
     wait_s=0
@@ -278,19 +664,41 @@ except MarksCorrupt as exc:
     # os.replace, and a 2 s cadence is plenty for a human clicking a radio.
     deadline=$(( $(date +%s) + wait_s ))
     while :; do
-      BOOTH_SRC="$(cd "$(dirname -- "$(readlink -f -- "$0")")/.." && pwd)" python3 -c '
+      # CAPTURED, not streamed. Printing inside the loop wrote one whole JSON
+      # document per poll, so `booth marks b --wait | jq` got several values
+      # concatenated and could parse none of them. The wait is a wait; the
+      # print is the result, and it happens once.
+      rc=0
+      out="$(BOOTH_SRC="$(cd "$(dirname -- "$(readlink -f -- "$0")")/.." && pwd)" python3 -c '
 import json, os, pathlib, sys
 sys.path.insert(0, os.environ["BOOTH_SRC"])
-from booth.marks import as_dict, marks_for, open_marks
+try:
-marks = marks_for(pathlib.Path(sys.argv[1]))
+    from booth.marks import as_dict, marks_for, open_marks, read_error
-print(json.dumps({"marks": [as_dict(m) for m in marks],
+    booth = pathlib.Path(sys.argv[1])
+    # Ask FIRST whether the file is readable. `marks_for` answers "no marks"
+    # for a damaged file, which is the right answer for a page and the wrong
+    # one for a session that wants to know whether its question survived.
+    broken = read_error(booth)
+    if broken:
+        print(f"booth: {broken}", file=sys.stderr)
+        sys.exit(3)
+    marks = marks_for(booth)
+    doc = json.dumps({"marks": [as_dict(m) for m in marks],
                       "open": [m.id for m in open_marks(marks)]},
-                 ensure_ascii=False, indent=2))
+                     ensure_ascii=False, indent=2)
-sys.exit(1 if open_marks(marks) else 0)
+except Exception as exc:
-' "$DATA/$name" && exit 0
+    print(f"booth: cannot read marks: {exc}", file=sys.stderr)
-      # exit 1 from the reader means at least one pick is still open
+    sys.exit(3)
-      if [ "$wait_s" -eq 0 ]; then exit 0; fi
+print(doc)
+sys.exit(2 if open_marks(marks) else 0)
+' "$DATA/$name")" || rc=$?
+      case "$rc" in
+        0) printf '%s\n' "$out"; exit 0 ;;                 # read ok, nothing open
+        2) if [ "$wait_s" -eq 0 ]; then printf '%s\n' "$out"; exit 0; fi ;;
+        *) echo "cannot read marks in $name" >&2; exit 3 ;;
+      esac
       if [ "$(date +%s)" -ge "$deadline" ]; then
+        printf '%s\n' "$out"
         echo "timed out after ${wait_s}s with marks still open in $name" >&2; exit 1
       fi
       sleep 2
@@ -304,20 +712,55 @@ sys.exit(1 if open_marks(marks) else 0)
     if [ "${1:-}" = "--wait" ]; then wait_s="${2:-3600}"; fi
     deadline=$(( $(date +%s) + wait_s ))
     while :; do
-      BOOTH_SRC="$(cd "$(dirname -- "$(readlink -f -- "$0")")/.." && pwd)" python3 -c '
+      rc=0
+      out="$(BOOTH_SRC="$(cd "$(dirname -- "$(readlink -f -- "$0")")/.." && pwd)" python3 -c '
 import json, os, pathlib, sys
 sys.path.insert(0, os.environ["BOOTH_SRC"])
-from booth.marks import marks_for
+try:
+    from booth.marks import marks_for, open_marks, read_error
     booth, mid = sys.argv[1:3]
-m = next((x for x in marks_for(pathlib.Path(booth)) if x.id == mid), None)
+    broken = read_error(pathlib.Path(booth))
+    if broken:
+        print(f"booth: {broken}", file=sys.stderr)
+        sys.exit(3)
+    # id AND shape, matching the web route. Matching on id alone reported a
+    # note id as "unanswered" and then polled it for an hour — a question that
+    # could never be answered because it was never a question.
+    marks = marks_for(pathlib.Path(booth))
+    m = next((x for x in marks if x.id == mid and x.shape == "pick"), None)
+    # THE openness predicate, not a second spelling of it. `answer is None` is
+    # what this read used to test, and it disagreed with `marks --wait` on a
+    # PARTIALLY answered pick: one verb returned the half-filled form while the
+    # other blocked on the same booth at the same instant. U2 put openness in
+    # one function precisely so the two could not drift.
+    still_open = m is not None and m in open_marks(marks)
+except Exception as exc:
+    print(f"booth: cannot read marks: {exc}", file=sys.stderr)
+    sys.exit(3)
 if m is None:
     sys.exit(2)
-if m.answer is None:
+if m.error:
+    # Not open, and never going to be: the web route refuses this form with a
+    # 400, so waiting on it is waiting on nothing. `marks --wait` already
+    # returns immediately here; this is the other half of that agreement.
+    print(f"booth: pick is broken and cannot be answered: {m.error}",
+          file=sys.stderr)
+    sys.exit(4)
+if still_open:
     sys.exit(1)
 print(json.dumps(m.answer, ensure_ascii=False, indent=2))
-' "$DATA/$name" "$mid" && exit 0
+' "$DATA/$name" "$mid")" || rc=$?
-      rc=$?
+      case "$rc" in
-      if [ "$rc" -eq 2 ]; then echo "no such pick: $name/$mid" >&2; exit 1; fi
+        0) printf '%s\n' "$out"; exit 0 ;;
+        2) echo "no such pick: $name/$mid" >&2; exit 2 ;;
+        # A read that FAILED is not "not yet". Conflating them sent --wait
+        # spinning for the full hour on a broken file and then blamed the
+        # operator for not answering.
+        3) echo "cannot read marks in $name" >&2; exit 3 ;;
+        # A pick that hydrated broken is refused by the web route, so no answer
+        # can ever land. Waiting on it is waiting on nothing.
+        4) exit 4 ;;
+      esac
       if [ "$wait_s" -eq 0 ]; then echo "unanswered: $URL/b/$name/#mark-$mid" >&2; exit 1; fi
       if [ "$(date +%s)" -ge "$deadline" ]; then
         echo "timed out after ${wait_s}s waiting on $name/$mid" >&2; exit 1

scripts/layout-probe.py

@@ -19,7 +19,22 @@ USAGE
     <a python with playwright> scripts/layout-probe.py [URL ...]
 
 Exits 0 if every control is hittable, 1 if any is occluded. No arguments
-probes the booth index and every booth linked from it.
+probes the INDEX ONLY — it does not follow booth links, and the docstring
+claimed it did until 2026-09-22. Pass booth URLs explicitly to cover them:
+
+    scripts/layout-probe.py http://10.100.10.50:8090/{,b/my-run/}
+
+⚠ PROBING A BOOTH PAGE RESETS THAT BOOTH'S TTL CLOCK (U4). A GET of `/b/<n>/`
+is a view, and a view is activity — that is the rule, and this script is not
+exempt from it just because it is ours. Sweeping every booth page therefore
+buys every booth another full TTL. Harmless and recoverable (nothing is
+deleted, things merely live longer), named here so nobody debugs it later as a
+sweeper that stopped working. The index-only default does NOT do this: browsing
+the index is deliberately not a view.
+
+⚠ In zsh an unquoted `$URLS` does NOT word-split, so a variable holding
+several URLs arrives as ONE argument and the probe silently reports
+"2 page(s)" while covering two. Use an array and `"${URLS[@]}"`.
 """
 import sys
 from playwright.sync_api import sync_playwright
@@ -62,6 +77,32 @@ def probe(page, url: str) -> list[str]:
             card.hover(timeout=1500)
         except Exception:
             pass
+    # ⚠ OPEN EVERY <details> FIRST. A control inside a CLOSED one is laid out
+    # but sits outside its collapsed parent's box, so `elementFromPoint` at its
+    # centre returns an ancestor and it reports OCCLUDED — 23 of them on
+    # `sindra-set`, every one a false positive, because the only way an operator
+    # reaches that button is by opening the disclosure first. Verified both
+    # ways: closed -> elementFromPoint returns div.gallery; opened -> the button
+    # itself, and a real trial click lands on it.
+    #
+    # Opening rather than SKIPPING is deliberate. Skipping would make the probe
+    # quiet by declaring put-away controls out of scope, and the add-note button
+    # inside `details.item-addnote` is exactly the kind of control this
+    # instrument exists to check. Open it and ask the real question.
+    #
+    # ⚠ ONE evaluate over the whole document, NOT a locator loop. `.all()` hands
+    # back positional locators that re-resolve against the CURRENT DOM, and
+    # `details:not([open])` stops matching an element the moment it is opened —
+    # so opening them one at a time shrinks the set underneath the indices and
+    # some are never opened at all. That left exactly the closed-<details>
+    # false positives this block exists to remove: 1 on booth-redesign, 3 on
+    # cr123a-to-d-sleeve, stable across five runs and invisible as a bug
+    # because a false positive looks like a finding. Measured both ways at
+    # 150 ms and 1000 ms settle: the loop reports them at either wait, the
+    # single pass reports none at either. The variable was the method, not the
+    # timing.
+    page.evaluate("document.querySelectorAll('details').forEach(d => d.open = true)")
+    page.wait_for_timeout(150)
     for el in page.locator("button, a.dl-link, a.thumb").all():
         try:
             # ⚠ elementFromPoint is VIEWPORT-relative. Without scrolling first,

tests/test_asks.py

@@ -13,7 +13,7 @@ import pathlib
 import pytest
 from fastapi.testclient import TestClient
 
-from booth.app import build_gallery, create_app, list_booths
+from booth.app import EMBED_SCRIPT_TAG, build_gallery, create_app, list_booths
 from booth.asks import (
     ANSWER_SUFFIX,
     ASK_SUFFIX,
@@ -410,32 +410,37 @@ def test_declare_pick_accepts_a_full_multi_doc(tmp_path):
 # into the verbatim page plus a standalone /asks page that carries the forms.
 
 
-def test_verbatim_booth_renders_the_ask_inline(client):
+def test_verbatim_booth_offers_the_ask_over_the_seam(client):
+    """U3: the report is served as written and the ask crosses the declared seam.
+    Before U3 the fragments were substituted into the page body by regex; the
+    guarantee that the ask is reachable FROM THE REPORT, not from another page,
+    is unchanged — it is the delivery that moved."""
     c, data = client
     b = _ask(data / "b")
     (b / "index.html").write_text("<!doctype html><title>report</title><body>hi</body>")
     html = c.get("/b/b/").text
     assert "hi" in html                          # the report is still served verbatim
-    assert "Which render wins?" in html          # ...with the ask ON it, not elsewhere
+    assert "Which render wins?" not in html      # ...and NOTHING was injected into it
-    assert 'type="radio"' in html and 'action="/b/b/answer"' in html
+    assert html.endswith(EMBED_SCRIPT_TAG)
-    assert "bk-ask" in html                      # self-contained fragment styles
+    (m,) = c.get("/b/b/embed.json").json()["marks"]
-    assert "booth-nav-asks" in html              # chip remains, as a jump link
+    assert "Which render wins?" in m["whole"]
-    assert "#bk-ask-winner-top" in html
+    assert 'type="radio"' in m["whole"] and 'action="/b/b/answer"' in m["submit"]
 
 
 def test_verbatim_chip_disappears_once_answered(client):
     c, data = client
     b = _ask(data / "b")
     (b / "index.html").write_text("<!doctype html><body>hi</body>")
+    assert c.get("/b/b/embed.json").json()["open"] == ["winner"]
     answer_pick(b, "winner", "A — baseline")
-    assert "booth-nav-asks" not in c.get("/b/b/").text
+    assert c.get("/b/b/embed.json").json()["open"] == []
 
 
 def test_verbatim_booth_without_asks_is_untouched(client):
     c, data = client
     (data / "b").mkdir()
     (data / "b" / "index.html").write_text("<!doctype html><body>hi</body>")
-    assert "booth-nav-asks" not in c.get("/b/b/").text
+    assert c.get("/b/b/embed.json").json()["marks"] == []
 
 
 def test_asks_page_renders_forms_and_answers_back_to_itself(client):
@@ -477,42 +482,64 @@ def test_asks_page_shows_a_single_ask_title(client):
     assert "emmie — pick the anchor" in c.get("/b/b/marks").text
 
 
-# ---- inline placement in a verbatim report -----------------------------------
+# ---- placement in a verbatim report ------------------------------------------
 #
 # Operator verdict 2026-09-09 on the separate /asks page: "the asks should be
 # inline with the artifacts, not on a separate page." A four-voice audition wants
 # each voice's radio group under that voice's audio, and one submit for the lot.
-
+#
+# U3 kept the semantics and moved the mechanism. The author still marks up where
+# each piece goes; the pieces are still rendered by the `_ask_inline.html`
+# macros; they now reach the page through `/b/<name>/embed.json` and are mounted
+# by `/_booth/embed.js` instead of substituted into the author's tags by regex.
+#
+# So the placement ASSERTIONS moved too, and where each half lives is not
+# arbitrary: what the server offers is checked here, in Python; where it LANDS,
+# and whether a form scattered down a report actually submits, is checked in
+# tests/test_embed_browser.py against a real DOM. No string assertion can see
+# the second thing, and that is exactly the part the operator depends on.
 
 REPORT = """<!doctype html><title>audition</title><body>
 <h1>Three voices</h1>
 <section id="lawson"><audio src="a.wav"></audio>
 <div data-booth-ask="batch:r1"></div></section>
 <section id="jo"><audio src="b.wav"></audio>
-<!-- booth:ask batch:r2 --></section>
+<div data-booth-mark="batch:r2"></div></section>
 <div data-booth-ask-submit="batch"></div>
+<script src="/_booth/embed.js" defer></script>
 </body>"""
 
 
-def test_per_question_placeholders_land_where_the_author_put_them(client):
+def test_the_author_markup_is_never_touched_by_the_server(client):
+    """The whole point of the seam. A page that declares it comes back exactly
+    as written — placeholders still empty, waiting for the DOM."""
     c, data = client
     b = _multi(data / "b")
     (b / "index.html").write_text(REPORT)
-    html = c.get("/b/b/").text
+    assert c.get("/b/b/").text == REPORT
-    # each group is inside its own section, in document order
-    lawson = html.index('id="lawson"')
-    jo = html.index('id="jo"')
-    assert lawson < html.index('name="choice.r1"') < jo
-    assert jo < html.index('name="choice.r2"')
-    # one shared form, bound by the HTML5 form= attribute, submitted once
-    assert html.count('<form id="bk-ask-form-batch"') == 1
-    assert html.count('action="/b/b/answer"') == 1
-    assert html.count('form="bk-ask-form-batch"') >= 4
-    # the submit block landed at its own placeholder, not appended after </body>
-    assert html.index("bk-ask-form-batch") < html.index("</body>")
 
 
-def test_inline_form_submits_every_question_in_one_post(client):
+def test_every_piece_the_author_can_place_is_offered(client):
+    """One fragment per addressable piece: the whole ask, each question, and the
+    submit block that carries the shared <form>. The author's markup decides
+    which are used; the payload never decides for them."""
+    c, data = client
+    b = _multi(data / "b")
+    (m,) = c.get("/b/b/embed.json").json()["marks"]
+    assert [q["key"] for q in m["questions"]] == ["r1", "r2"]
+    assert 'name="choice.r1"' in m["questions"][0]["html"]
+    assert 'name="choice.r2"' in m["questions"][1]["html"]
+    # ONE form, and it lives with the submit block, so question groups scattered
+    # down a report bind to it by id from wherever they sit.
+    assert m["submit"].count('<form id="bk-ask-form-batch"') == 1
+    assert m["submit"].count('action="/b/b/answer"') == 1
+    assert 'form="bk-ask-form-batch"' in m["questions"][0]["html"]
+    assert 'form="bk-ask-form-batch"' in m["questions"][1]["html"]
+
+
+def test_a_scattered_form_still_posts_as_one_answer(client):
+    """The POST half of the multi-question guarantee, which U3 did not touch:
+    every question in one request, or the route refuses it."""
     c, data = client
     b = _multi(data / "b")
     (b / "index.html").write_text(REPORT)
@@ -521,44 +548,21 @@ def test_inline_form_submits_every_question_in_one_post(client):
     assert r.status_code == 303
     ans = _answer_of(b, "batch")
     assert ans["answers"]["r1"]["choice"] == "keep" and ans["answers"]["r2"]["choice"] == "d"
-    # and the recorded pick now shows inline, on the report itself
+    # and the recorded pick comes back marked answered, on the report's own seam
-    html = c.get("/b/b/").text
+    (m,) = c.get("/b/b/embed.json").json()["marks"]
-    assert "recorded:" in html and "bk-done" in html
+    assert "recorded:" in m["whole"] and "bk-done" in m["whole"]
-    assert 'value="keep" required  checked' in html.replace("\n", " ") or "checked" in html
+    assert "checked" in m["questions"][0]["html"]
 
 
-def test_whole_ask_placeholder_renders_everything_there(client):
+def test_the_page_carries_no_fragment_styles(client):
-    c, data = client
+    """`styles()` is gone from the template: the scoped `.bk-ask-*` rules live in
-    b = _ask(data / "b")
+    embed.js, next to the code that mounts them. One asset, emitted once by
-    (b / "index.html").write_text('<!doctype html><body><p>x</p><div data-booth-ask="winner"></div></body>')
+    construction rather than by a seen-set."""
-    html = c.get("/b/b/").text
-    assert html.index("Which render wins?") > html.index("<p>x</p>")
-    assert html.index("bk-ask-go") < html.index("</body>")     # submit placed inline too
-
-
-def test_placeholder_for_a_missing_ask_is_left_alone(client):
-    c, data = client
-    b = _ask(data / "b")
-    (b / "index.html").write_text('<!doctype html><body><div data-booth-ask="typo"></div></body>')
-    html = c.get("/b/b/").text
-    assert 'data-booth-ask="typo"' in html          # author's markup untouched, not blanked
-    assert "Which render wins?" in html             # the real ask still appended, never lost
-
-
-def test_questions_placed_without_a_submit_still_get_one(client):
-    c, data = client
-    b = _multi(data / "b")
-    (b / "index.html").write_text('<!doctype html><body><div data-booth-ask="batch:r1"></div></body>')
-    html = c.get("/b/b/").text
-    assert html.count('<form id="bk-ask-form-batch"') == 1     # appended, so it is submittable
-    assert 'name="choice.r2"' in html                          # r2 unplaced -> must still appear
-
-
-def test_styles_are_emitted_once(client):
     c, data = client
     b = _multi(data / "b")
     (b / "index.html").write_text(REPORT)
-    assert c.get("/b/b/").text.count(".bk-ask-opt:has(input:checked)") == 1
+    assert ".bk-ask-opt:has(input:checked)" not in c.get("/b/b/").text
+    assert c.get("/_booth/embed.js").text.count(".bk-ask-opt:has(input:checked)") == 1
 
 
 def test_radios_are_not_html_required_anywhere(client):
@@ -566,20 +570,20 @@ def test_radios_are_not_html_required_anywhere(client):
     is exactly what stopped the operator leaving one blank."""
     c, data = client
     b = _multi(data / "b")
-    assert "required" not in c.get("/b/b/").text
+    (m,) = c.get("/b/b/embed.json").json()["marks"]
-    (b / "index.html").write_text('<!doctype html><body><div data-booth-ask="batch"></div></body>')
+    assert "required" not in m["whole"]
-    assert "required" not in c.get("/b/b/").text
+    assert not any("required" in q["html"] for q in m["questions"])
     assert "required" not in c.get("/b/b/marks").text
 
 
-def test_partial_answer_renders_as_skipped_inline(client):
+def test_partial_answer_renders_as_skipped(client):
     c, data = client
     b = _multi(data / "b")
-    (b / "index.html").write_text('<!doctype html><body><div data-booth-ask="batch"></div></body>')
+    (b / "index.html").write_text(REPORT)
     c.post("/b/b/answer", data={"ask": "batch", "choice.r1": "keep"})
-    html = c.get("/b/b/").text
+    (m,) = c.get("/b/b/embed.json").json()["marks"]
-    assert "bk-skip" in html and "left blank" in html
+    assert "bk-skip" in m["whole"] and "left blank" in m["whole"]
-    assert "1 of 2 answered" in html
+    assert "1 of 2 answered" in m["submit"]
 
 
 def test_empty_submission_is_refused_with_400(client):

tests/test_benches.py

@@ -0,0 +1,889 @@
+"""U6 — benches: a running thing, registered.
+
+The contract is docs/contracts/u6_benches.contract.md. Every test here names
+the invariant it falsifies, and each is written to go RED under the change that
+defeats that invariant — not merely to assert the outcome the author had in
+mind. (The U4 round shipped seven falsifiers of which five stayed green under
+the very change they forbade; see persistent-memory.d/2026-09-22-vacuous-falsifiers.md.)
+"""
+
+from __future__ import annotations
+
+import ast
+import json
+import os
+import pathlib
+import sys
+
+import pytest
+
+sys.path.insert(0, str(pathlib.Path(__file__).parent.parent))
+
+from booth.benches import (  # noqa: E402
+    BENCHES_FILE,
+    BENCH_STATES,
+    Bench,
+    normalize_bench_url,
+    order_benches,
+    read_benches,
+    remove_bench,
+    set_bench_state,
+    upsert_bench,
+)
+from booth.links import booth_target  # noqa: E402
+
+
+# ---- INV-6: the identity collapses a re-post and NOTHING else ---------------
+#
+# Both directions from ONE fixture. A test that only checked the talk collapse
+# would pass under origin normalization, which is the measurably wrong rule:
+# on the live board it merges eight distinct gitea repositories into one row.
+
+# Measured on the live board, 2026-09-22.
+GITEA_EIGHT = [
+    "https://gitea.phasefinal.com/vh/bifrost/issues/17",
+    "https://gitea.phasefinal.com/vh/brokkr-smithy/src/commit/6adcde6/research/landscape-scans/open-weight-releases-2026-09-15.md",
+    "https://gitea.phasefinal.com/vh/cicada",
+    "https://gitea.phasefinal.com/vh/draupnir",
+    "https://gitea.phasefinal.com/vh/-/packages/pypi/bifrost/1.2.0",
+    "https://gitea.phasefinal.com/vh/-/packages/pypi/bifrost/1.2.1",
+    "https://gitea.phasefinal.com/vh/peedlar",
+    "https://gitea.phasefinal.com/vh/peedlar/releases/tag/v0.3.0",
+]
+TALK_FIVE = ["https://talk.nh3.phasefinal.com:8092/"] * 5
+
+
+def test_eight_distinct_repos_stay_eight(tmp_path):
+    """INV-6, the direction origin-normalization gets WRONG. Defeating change:
+    normalizing to scheme://host:port. This goes red under it; the collapse
+    test below does not."""
+    for i, u in enumerate(GITEA_EIGHT):
+        upsert_bench(tmp_path, u, f"repo {i}", "vh")
+    benches, err = read_benches(tmp_path)
+    assert err is None
+    assert len(benches) == 8, [b.id for b in benches]
+
+
+def test_five_reposts_of_one_bench_collapse(tmp_path):
+    """INV-6, the direction the IA doc names. `talk` is on the live board five
+    times; the registry must hold one row, carrying the LAST name."""
+    for i, u in enumerate(TALK_FIVE):
+        _, created = upsert_bench(tmp_path, u, f"talk v{i}", "nh3-dev")
+        assert created is (i == 0)
+    benches, _ = read_benches(tmp_path)
+    assert len(benches) == 1
+    assert benches[0].name == "talk v4"
+
+
+def test_two_lrpg_surfaces_on_one_origin_stay_two(tmp_path):
+    """INV-6. The IA doc's OWN example of two real benches shares an origin."""
+    upsert_bench(tmp_path, "http://10.100.10.50:8321/Authoring%20Studio.dc.html", "authoring", "ldp-dev")
+    upsert_bench(tmp_path, "http://10.100.10.50:8321/GM%20Playback.dc.html", "gm", "ldp-dev")
+    assert len(read_benches(tmp_path)[0]) == 2
+
+
+def test_query_is_part_of_the_identity(tmp_path):
+    """INV-6. Three ShutterChute rows differ ONLY by `?token=`; they are three
+    links, not one bench posted three times. Defeating change: dropping query."""
+    base = "http://10.100.10.50:8477/?token="
+    for tok in ("aaa", "bbb", "ccc"):
+        upsert_bench(tmp_path, base + tok, "shutterchute", "nh3-dev")
+    assert len(read_benches(tmp_path)[0]) == 3
+
+
+@pytest.mark.parametrize("a,b", [
+    ("http://x.test/", "http://X.TEST"),             # host case + bare-slash path
+    ("http://x.test:80/p", "http://x.test/p"),        # default port
+    ("https://x.test:443/p", "https://x.test/p"),
+    ("http://x.test/p#frag", "http://x.test/p"),      # fragment dropped
+    ("  http://x.test/p  ", "http://x.test/p"),       # whitespace
+    ("http://[::1]:80/a", "http://[::1]/a"),          # default port, bracketed
+    ("http://[::1]/A", "http://[::1]/A"),             # bracket round-trips
+])
+def test_these_pairs_are_one_bench(a, b):
+    """INV-6. Each pair is the SAME resource reached two ways."""
+    assert normalize_bench_url(a) == normalize_bench_url(b)
+
+
+@pytest.mark.parametrize("a,b", [
+    ("http://x.test/p", "http://x.test/p/"),          # trailing slash on a REAL path
+    ("http://x.test/p", "http://x.test/P"),           # path case
+    ("http://x.test/?a=1&b=2", "http://x.test/?b=2&a=1"),  # query order is opaque
+    ("http://x.test:8092/", "https://x.test:8092/"),  # scheme
+    # A NON-DEFAULT PORT IS PART OF THE IDENTITY. Without this vector, "always
+    # omit the port" passes every other row in this file — caught by the cold
+    # panel's per-invariant "what would still pass" pass, not by us.
+    ("http://x.test:8092/p", "http://x.test/p"),
+    ("https://x.test:8443/p", "https://x.test/p"),
+])
+def test_these_pairs_are_two_benches(a, b):
+    """INV-6, the other direction. Each pair MAY be two different resources, and
+    the registry must not decide otherwise on the operator's behalf."""
+    assert normalize_bench_url(a) != normalize_bench_url(b)
+
+
+@pytest.mark.parametrize("bad", [
+    "", "   ", "not a url", "ftp://x.test/f", "file:///etc/passwd",
+    "http://", "https:///path", "//x.test/p", "javascript:alert(1)",
+])
+def test_refused_urls_raise_with_a_reason(bad):
+    with pytest.raises(ValueError) as e:
+        normalize_bench_url(bad)
+    assert str(e.value).strip(), "a refusal with no reason is a refusal the CLI cannot print"
+
+
+def test_credentials_are_refused_not_stripped():
+    """Stripping would register a bench whose URL no longer works while telling
+    the poster it succeeded — and put a credential on an unauthenticated LAN
+    surface on the way. Defeating change: `netloc.rpartition('@')[2]`."""
+    with pytest.raises(ValueError):
+        normalize_bench_url("https://user:hunter2@x.test/p")
+
+
+# ---- INV-7: `url` is what a click goes to; `id` is never the href -----------
+
+
+def test_the_stored_url_is_the_raw_string(tmp_path):
+    """INV-7. Defeating change: storing the normalized form as `url` because it
+    is 'the clean one'. Every field that differs is asserted, byte for byte."""
+    raw = "  HTTP://X.Test:80/Some%20Path/?b=2&a=1#frag  "
+    bench, _ = upsert_bench(tmp_path, raw, "n", "o")
+    assert bench.url == raw.strip()
+    assert bench.id != bench.url
+    assert bench.id == "http://x.test/Some%20Path/?b=2&a=1"
+    assert read_benches(tmp_path)[0][0].url == raw.strip()
+
+
+# ---- INV-4: the rendered order is TOTAL and stated --------------------------
+
+
+def test_same_name_benches_do_not_swap():
+    """INV-4. Defeating change: dropping the `id` tie-break.
+
+    THIS TEST USED TO GO THROUGH THE REGISTRY AND COULD NOT FAIL. `_write_all`
+    serializes with `sort_keys=True`, so whatever order two benches were
+    inserted in, they came back off disk already id-sorted — and removing the
+    tie-break from `order_benches` left it green. A vacuous falsifier of
+    exactly the shape persistent-memory.d/2026-09-22-vacuous-falsifiers.md
+    describes: it asserted the outcome the author had in mind rather than the
+    discriminator the invariant names. Caught by the cold panel (hulda, solo),
+    not by us.
+
+    So it calls `order_benches` DIRECTLY, with records that tie on both prior
+    keys, presented in both orders. Nothing upstream can pre-sort them.
+    """
+    def recs(order):
+        pair = [
+            Bench(id="http://a.test/", url="http://a.test/", name="same name",
+                  owner="o", state="live", added="", updated=""),
+            Bench(id="http://b.test/", url="http://b.test/", name="same name",
+                  owner="o", state="live", added="", updated=""),
+        ]
+        return pair if order else list(reversed(pair))
+    assert [b.id for b in order_benches(recs(0))] == \
+           [b.id for b in order_benches(recs(1))]
+    assert [b.id for b in order_benches(recs(1))] == ["http://a.test/", "http://b.test/"]
+
+
+def test_state_ranks_before_name(tmp_path):
+    """INV-4. live → promoted → retired, THEN name. Defeating change: ordering
+    by name alone, which a fixture of three same-state benches cannot see."""
+    upsert_bench(tmp_path, "http://a.test/", "aaa", "o")   # would sort first by name
+    upsert_bench(tmp_path, "http://z.test/", "zzz", "o")
+    set_bench_state(tmp_path, normalize_bench_url("http://a.test/"), "retired")
+    assert [b.name for b in read_benches(tmp_path)[0]] == ["zzz", "aaa"]
+
+
+def test_order_is_case_insensitive_on_name(tmp_path):
+    upsert_bench(tmp_path, "http://b.test/", "Bravo", "o")
+    upsert_bench(tmp_path, "http://a.test/", "alpha", "o")
+    assert [b.name for b in read_benches(tmp_path)[0]] == ["alpha", "Bravo"]
+
+
+def test_order_benches_is_pure(tmp_path):
+    """INV-4. Defeating change: `order_benches` doing I/O or sorting in place.
+    Called with records belonging to NO root, it must still answer."""
+    made = [Bench(id=f"http://{c}.test/", url=f"http://{c}.test/", name=c,
+                  owner="o", state="live", added="", updated="") for c in "ba"]
+    assert [b.name for b in order_benches(made)] == ["a", "b"]
+    assert [b.name for b in made] == ["b", "a"], "input was mutated"
+
+
+# ---- INV-5: the read cannot raise, and cannot cost the caller unboundedly ---
+
+
+def _write_raw(root: pathlib.Path, payload: str) -> None:
+    (root / BENCHES_FILE).write_text(payload)
+
+
+def test_absent_registry_is_not_an_error(tmp_path):
+    benches, err = read_benches(tmp_path)
+    assert benches == [] and err is None
+
+
+@pytest.mark.parametrize("payload,label", [
+    ("this is not json", "non-JSON bytes"),
+    ("[]", "valid JSON of the wrong top-level shape"),
+    ('{"benches": []}', "the list shape this unit deliberately does not use"),
+    ('{"http://a/": "a string, not a record"}', "a value of the wrong type"),
+    ('{"http://a/": {"name": [], "owner": "o", "state": "live"}}', "a FIELD of the wrong type"),
+    ('{"http://a/": {"name": "n", "owner": "o", "state": "invented"}}', "an unknown state"),
+])
+def test_damaged_registries_report_rather_than_raise(tmp_path, payload, label):
+    """INV-5. Defeating change: `json.load` with no guard, or `except: pass`
+    which would report absent. The error must be NON-EMPTY — 'damaged' and
+    'absent' must not render the same, because only one of them needs a human.
+    The wrong-typed-FIELD row is the shape currently 500ing the gallery
+    elsewhere in this service."""
+    _write_raw(tmp_path, payload)
+    benches, err = read_benches(tmp_path)
+    assert err, f"{label} reported no error"
+    assert benches == []
+
+
+def test_oversized_registry_is_refused_by_size_before_parsing(tmp_path):
+    """INV-5. Defeating change: parsing first and checking length after, which
+    costs the caller the whole file. A FIFO has st_size 0, so the guard must
+    bound the READ, not trust the stat — the 2026-09-22 hang lesson."""
+    import booth.benches as B
+    _write_raw(tmp_path, '{"http://a/": {"name": "' + "x" * B.BENCHES_MAX_BYTES + '"}}')
+    benches, err = read_benches(tmp_path)
+    assert err and benches == []
+    # AND PROVE THE PARSE WAS NEVER REACHED. Asserting only the eventual result
+    # passes an implementation that loads the whole document and checks its
+    # length afterwards — which costs the caller exactly what the cap exists to
+    # save. Booby-trap json.loads: if it runs, the test says so. Cold panel,
+    # hulda F11.
+    import json as _json
+    tripped = []
+    real = _json.loads
+
+    def trap(*a, **k):
+        tripped.append(True)
+        return real(*a, **k)
+    B.json.loads = trap
+    try:
+        benches, err = read_benches(tmp_path)
+    finally:
+        B.json.loads = real
+    assert err and not tripped, "the oversized registry was parsed before it was refused"
+
+
+@pytest.mark.skipif(os.geteuid() == 0, reason="root ignores the mode bit")
+def test_unreadable_registry_reports_rather_than_raises(tmp_path):
+    p = tmp_path / BENCHES_FILE
+    p.write_text("{}")
+    p.chmod(0o000)
+    try:
+        benches, err = read_benches(tmp_path)
+        assert err and benches == []
+    finally:
+        p.chmod(0o644)
+
+
+# ---- INV-1: one module knows the registry's filename ------------------------
+
+
+def test_only_benches_py_names_the_registry_file():
+    """INV-1. Defeating change: a route reading `.benches.json` directly to save
+    an import. Asserting that the panel renders would pass under exactly that."""
+    root = pathlib.Path(__file__).parent.parent
+    offenders = []
+    for f in list((root / "booth").rglob("*.py")) + [root / "scripts" / "booth"]:
+        if f.name == "benches.py":
+            continue
+        if ".benches.json" in f.read_text():
+            offenders.append(str(f.relative_to(root)))
+    assert not offenders, f"the registry filename is hard-coded outside benches.py: {offenders}"
+
+
+# ---- INV-9: stdlib-only, AND sibling-free (seam review SR-1) ----------------
+
+
+def test_benches_is_stdlib_only_and_imports_no_sibling():
+    """INV-9. The PARAMETRIZED test in test_marks.py exempts `booth` on purpose,
+    so it cannot catch `from booth.links import booth_target` — which is exactly
+    the import this unit tempts an implementer into. This is the strict copy,
+    mirroring tests/test_manifest.py. Seam review SR-1."""
+    src = pathlib.Path(__file__).parent.parent / "booth" / "benches.py"
+    tree = ast.parse(src.read_text())
+    roots = set()
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Import):
+            roots.update(a.name.split(".")[0] for a in node.names)
+        elif isinstance(node, ast.ImportFrom):
+            roots.add("booth" if node.level else (node.module or "").split(".")[0])
+    outside = {r for r in roots if r and r not in sys.stdlib_module_names}
+    assert not outside, f"booth/benches.py imports outside the stdlib (booth.* included): {sorted(outside)}"
+    # A STRING IMPORT IS INVISIBLE TO THE WALK ABOVE. `__import__("booth.links")`
+    # or `importlib.import_module(...)` inside a function defeats it entirely,
+    # and that is the exact shape someone reaches for when a sibling import is
+    # refused by review. Caught by the cold panel's per-invariant vacuity pass.
+    called = {n.func.id for n in ast.walk(tree)
+              if isinstance(n, ast.Call) and isinstance(n.func, ast.Name)}
+    assert "__import__" not in called, "benches.py imports by string, defeating the AST walk"
+    assert "importlib" not in roots, "benches.py can import anything at runtime via importlib"
+
+
+# ---- INV-2: ONE predicate decides what a booth URL is -----------------------
+
+# Every row is (url, expected booth name or None). Run against BOTH callers.
+BOOTH_URL_TABLE = [
+    ("http://10.100.10.50:8090/b/sindra-bakeoff/", "sindra-bakeoff"),
+    ("http://10.100.10.50:8090/b/sindra-bakeoff", "sindra-bakeoff"),
+    ("http://localhost:8090/b/x/", "x"),
+    ("http://NH3-DEV.nh3.internal:8090/b/x/", "x"),      # host-agnostic, any case
+    ("https://10.100.10.50:8090/b/x/", "x"),             # scheme-agnostic
+    ("http://10.100.10.50:8090/b/my%20booth/", "my booth"),   # SR-7: decoded
+    ("http://10.100.10.50:8090/b/x/zoom/a.png", "x"),    # nested path
+    ("http://10.100.10.50:8090/b/x/?q=1", "x"),          # query
+    ("http://10.100.10.50:8090/b/x/#frag", "x"),
+    ("http://10.100.10.50:8090/", None),                 # the Booth root IS a bench
+    ("http://10.100.10.50:8090/b/", None),               # no name
+    ("http://10.100.10.50:8090/b//", None),
+    ("http://10.100.10.50:8090/b/.hidden/", None),       # resolve_booth's rules
+    ("http://10.100.10.50:8090/b/%2e%2e/", None),        # decoded `..`
+    ("http://10.100.10.50:8090/b/a%2Fb/", None),         # decoded separator
+    ("https://gitea.phasefinal.com/vh/peedlar", None),
+    ("not a url at all", None),
+    # THE ACCEPTED COST, MADE EXPLICIT. The predicate is host-agnostic on
+    # purpose — a host allowlist fails OPEN on whichever name somebody reaches
+    # this service by next — so a third-party URL with a `/b/<x>` path reads as
+    # a booth link and is refused. The contract names this trade-off; the table
+    # had no row exercising it, so nothing pinned the behaviour either way.
+    # Cold panel, hulda F10.
+    ("https://example.com/b/not-ours/", "not-ours"),
+]
+
+
+@pytest.mark.parametrize("url,expected", BOOTH_URL_TABLE)
+def test_booth_target_classifies(url, expected):
+    """INV-2. The table is shared with the CLI refusal test and the dead-marker
+    test, so a second implementation in either place goes red here or there."""
+    assert booth_target(url) == expected
+
+
+def test_booth_target_never_raises():
+    """A board row is arbitrary operator-editable text; a predicate that raises
+    on one row takes the whole page. Defeating change: `urlsplit` unguarded."""
+    for junk in ["", "   ", "http://[oops", "\x00", "://", "http://]"]:
+        assert booth_target(junk) is None
+
+
+# ---- upsert semantics -------------------------------------------------------
+
+
+def test_added_survives_reregistration_updated_does_not(tmp_path):
+    first, created = upsert_bench(tmp_path, "http://a.test/", "one", "o1")
+    assert created
+    second, created = upsert_bench(tmp_path, "http://a.test/", "two", "o2")
+    assert not created
+    assert second.added == first.added
+    assert second.name == "two" and second.owner == "o2"
+    assert second.updated >= first.updated
+
+
+def test_state_survives_reregistration(tmp_path):
+    """A promoted bench that re-announces itself is still promoted — otherwise
+    every deploy silently demotes it."""
+    upsert_bench(tmp_path, "http://a.test/", "one", "o")
+    set_bench_state(tmp_path, normalize_bench_url("http://a.test/"), "promoted")
+    again, _ = upsert_bench(tmp_path, "http://a.test/", "one again", "o")
+    assert again.state == "promoted"
+
+
+def test_a_new_bench_is_live(tmp_path):
+    bench, _ = upsert_bench(tmp_path, "http://a.test/", "one", "o")
+    assert bench.state == "live" and bench.state in BENCH_STATES
+
+
+def test_set_state_refuses_an_unknown_state(tmp_path):
+    upsert_bench(tmp_path, "http://a.test/", "one", "o")
+    with pytest.raises(ValueError):
+        set_bench_state(tmp_path, normalize_bench_url("http://a.test/"), "invented")
+
+
+def test_set_state_and_remove_miss_cleanly(tmp_path):
+    assert set_bench_state(tmp_path, "http://nope/", "live") is None
+    assert remove_bench(tmp_path, "http://nope/") is None
+
+
+def test_remove_returns_the_record_and_drops_it(tmp_path):
+    upsert_bench(tmp_path, "http://a.test/", "one", "o")
+    gone = remove_bench(tmp_path, normalize_bench_url("http://a.test/"))
+    assert gone is not None and gone.name == "one"
+    assert read_benches(tmp_path)[0] == []
+
+
+def test_fields_are_capped_at_the_write(tmp_path):
+    from booth.benches import NAME_MAX, OWNER_MAX
+    bench, _ = upsert_bench(tmp_path, "http://a.test/", "n" * 500, "o" * 500)
+    assert len(bench.name) == NAME_MAX and len(bench.owner) == OWNER_MAX
+
+
+def test_a_write_over_a_damaged_registry_does_not_destroy_it(tmp_path):
+    """The 2026-09-21 lesson, in this unit's storage: reads are lenient, writes
+    are STRICT. A damaged registry must not be silently replaced by a fresh one
+    carrying only the new row — that is the marks-wipe bug in a new file."""
+    _write_raw(tmp_path, '{"http://a/": {"name": "real", "owner": "o", "state": "live"}, BROKEN')
+    before = (tmp_path / BENCHES_FILE).read_text()
+    with pytest.raises(ValueError):
+        upsert_bench(tmp_path, "http://b.test/", "new", "o")
+    assert (tmp_path / BENCHES_FILE).read_text() == before
+
+
+def test_the_on_disk_shape_is_an_object_keyed_by_id(tmp_path):
+    """Two rows with one identity are then impossible BY CONSTRUCTION rather
+    than by an upsert remembering to check."""
+    upsert_bench(tmp_path, "http://a.test/", "one", "o")
+    raw = json.loads((tmp_path / BENCHES_FILE).read_text())
+    # The key is the NORMALIZED url, so the bare "/" is already gone — which is
+    # the rule `test_these_pairs_are_one_bench` pins independently.
+    assert isinstance(raw, dict) and list(raw) == ["http://a.test"]
+    assert "id" not in raw["http://a.test"], "the key IS the id; storing it twice invites drift"
+
+
+# ---- the rendered surface ---------------------------------------------------
+#
+# The benches panel and the dead-row marker both live on the standing board's
+# page — the one booth carrying a links.md.
+
+from fastapi.testclient import TestClient  # noqa: E402
+
+from booth.app import create_app  # noqa: E402
+
+
+def _board(root: pathlib.Path, rows: str) -> pathlib.Path:
+    b = root / "links"
+    b.mkdir(parents=True, exist_ok=True)
+    (b / "links.md").write_text(rows)
+    return b
+
+
+def _client(root):
+    return TestClient(create_app(root, ttl_hours=24, start_sweeper=False))
+
+
+ROW_LIVE = "- [still here](http://10.100.10.50:8090/b/alive/) <sub>· x · 2026-09-01 00:00</sub>\n"
+ROW_DEAD = "- [swept](http://10.100.10.50:8090/b/gone/) <sub>· x · 2026-09-01 00:00</sub>\n"
+ROW_REF = "- [a repo](https://gitea.phasefinal.com/vh/peedlar) <sub>· x · 2026-09-01 00:00</sub>\n"
+
+
+def _dead_rows(body: str) -> list[str]:
+    """Board ROWS carrying the dead class.
+
+    Scoped to `<div class="board-row ...">` on purpose: the class name also
+    appears in base.html's stylesheet, so a whole-document substring test is
+    always true and can never go red — a vacuous falsifier of exactly the shape
+    persistent-memory.d/2026-09-22-vacuous-falsifiers.md describes.
+    """
+    return [ln for ln in body.splitlines()
+            if 'class="board-row' in ln and "board-dead" in ln]
+
+
+def test_a_dead_row_is_marked_and_a_live_one_is_not(tmp_path):
+    """The marker. Defeating change: marking every `/b/` row dead, or none.
+    Both a live target and a dead one are in ONE fixture, so a marker that is
+    constant in either direction goes red."""
+    (tmp_path / "alive").mkdir()
+    _board(tmp_path, ROW_LIVE + ROW_DEAD + ROW_REF)
+    r = _client(tmp_path).get("/b/links/")
+    assert r.status_code == 200
+    rows = _dead_rows(r.text)
+    assert len(rows) == 1, f"exactly one of the three rows is dead, got {rows}"
+    assert "/b/gone/" in r.text and "booth is gone" in r.text
+
+
+def test_the_marker_never_takes_the_page(tmp_path):
+    """Seam review SR-2. `resolve_booth` RAISES HTTPException(404); calling it
+    per row would turn one swept booth into a 404 for the whole board. This is
+    the test that goes red under that exact implementation."""
+    _board(tmp_path, ROW_DEAD * 5)
+    assert _client(tmp_path).get("/b/links/").status_code == 200
+
+
+def test_a_percent_encoded_booth_is_not_marked_dead(tmp_path):
+    """Seam review SR-7. `quote(name, safe="")` is how the service emits these,
+    so the marker must decode before it looks on disk. Defeating change:
+    comparing the raw path segment — which marks this row dead forever."""
+    (tmp_path / "my booth").mkdir()
+    _board(tmp_path, "- [x](http://10.100.10.50:8090/b/my%20booth/) <sub>· x · 2026-09-01 00:00</sub>\n")
+    assert _dead_rows(_client(tmp_path).get("/b/links/").text) == []
+
+
+def test_a_reference_row_is_never_marked_dead(tmp_path):
+    _board(tmp_path, ROW_REF)
+    assert _dead_rows(_client(tmp_path).get("/b/links/").text) == []
+
+
+def test_the_panel_renders_registered_benches(tmp_path):
+    _board(tmp_path, ROW_REF)
+    upsert_bench(tmp_path, "https://talk.nh3.phasefinal.com:8092/", "talk", "tts-dev")
+    body = _client(tmp_path).get("/b/links/").text
+    assert "talk" in body and "tts-dev" in body
+
+
+def test_the_anchor_href_is_the_raw_url_not_the_id(tmp_path):
+    """INV-7. Defeating change: rendering `bench.id` in the href because it is
+    'the clean one'. The raw URL here normalizes differently in three ways."""
+    raw = "HTTP://Talk.NH3.test:80/Some%20Path/?b=2&a=1#frag"
+    _board(tmp_path, ROW_REF)
+    upsert_bench(tmp_path, raw, "talk", "o")
+    body = _client(tmp_path).get("/b/links/").text
+    assert 'href="HTTP://Talk.NH3.test:80/Some%20Path/?b=2&amp;a=1#frag"' in body, \
+        "the href must be the URL as posted, byte for byte"
+
+
+@pytest.mark.parametrize("payload,label", [
+    (None, "absent"),
+    ("not json", "non-JSON"),
+    ("[]", "wrong top-level shape"),
+    ('{"http://a/": {"name": [], "owner": "o", "state": "live"}}', "wrong-typed field"),
+    ('{"http://a/": {"name": "n", "owner": "o", "state": "invented"}}', "unknown state"),
+    ("OVERSIZED", "over the size cap"),
+    ("UNREADABLE", "chmod 000"),
+    ("FIFO", "a named pipe"),
+])
+def test_a_damaged_registry_costs_its_panel_and_never_the_page(tmp_path, payload, label):
+    """INV-5, at the render. The v0.2.2 lesson: a poisoned sidecar returned 500
+    for `/` and `/healthz` across all 25 booths. The wrong-typed-FIELD row is
+    the shape currently 500ing the gallery elsewhere in this service, so it is
+    the one that matters most."""
+    _board(tmp_path, ROW_REF)
+    reg = tmp_path / BENCHES_FILE
+    if payload == "OVERSIZED":
+        from booth.benches import BENCHES_MAX_BYTES
+        reg.write_text('{"http://a/": {"name": "' + "x" * BENCHES_MAX_BYTES + '"}}')
+    elif payload == "UNREADABLE":
+        if os.geteuid() == 0:
+            pytest.skip("root ignores the mode bit")
+        reg.write_text("{}")
+        reg.chmod(0o000)
+    elif payload == "FIFO":
+        os.mkfifo(reg)
+    elif payload is not None:
+        reg.write_text(payload)
+    try:
+        c = _client(tmp_path)
+        body = c.get("/b/links/")
+        assert body.status_code == 200, label
+        assert c.get("/").status_code == 200, label
+        assert c.get("/healthz").status_code == 200, label
+        # AND THE ERROR IS VISIBLE. Asserting only 200 was the gap: a render
+        # that swallowed the failure and drew an empty panel passed every case
+        # here while telling the operator nothing needed fixing. Absent is the
+        # one case that must NOT show an error.
+        shown = "the bench registry could not be read" in body.text
+        assert shown is (payload is not None), label
+    finally:
+        if payload == "UNREADABLE":
+            reg.chmod(0o644)
+
+
+def test_damaged_and_absent_render_different_text(tmp_path):
+    """INV-5. Only ONE of them needs a human. Defeating change: `except: pass`
+    returning ([], None), which renders damaged exactly like absent."""
+    _board(tmp_path, ROW_REF)
+    absent = _client(tmp_path).get("/b/links/").text
+    (tmp_path / BENCHES_FILE).write_text("not json")
+    damaged = _client(tmp_path).get("/b/links/").text
+    assert absent != damaged
+
+
+def test_the_panel_does_not_render_on_an_ordinary_booth(tmp_path):
+    """A bench registry on every gallery page would be noise, and would cost a
+    read per booth page view for a surface that belongs to exactly one."""
+    (tmp_path / "ordinary").mkdir()
+    (tmp_path / "ordinary" / "a.png").write_bytes(b"\x89PNG\r\n\x1a\n")
+    upsert_bench(tmp_path, "https://talk.test/", "talk", "o")
+    assert "talk" not in _client(tmp_path).get("/b/ordinary/").text
+
+
+def test_the_benches_routes_round_trip(tmp_path):
+    _board(tmp_path, ROW_REF)
+    c = _client(tmp_path)
+    assert c.post("/b/links/bench-add", data={"url": "https://x.test/", "name": "ex"},
+                  follow_redirects=False).status_code in (302, 303)
+    assert "ex" in c.get("/b/links/").text
+    bid = normalize_bench_url("https://x.test/")
+    c.post("/b/links/bench-state", data={"bench": bid, "state": "retired"},
+           follow_redirects=False)
+    assert read_benches(tmp_path)[0][0].state == "retired"
+    c.post("/b/links/bench-remove", data={"bench": bid}, follow_redirects=False)
+    assert read_benches(tmp_path)[0] == []
+
+
+def test_a_bad_url_posted_to_the_route_does_not_500(tmp_path):
+    _board(tmp_path, ROW_REF)
+    c = _client(tmp_path)
+    r = c.post("/b/links/bench-add", data={"url": "ftp://x.test/f", "name": "ex"},
+               follow_redirects=False)
+    assert r.status_code in (302, 303, 400)
+    assert c.get("/b/links/").status_code == 200
+
+
+# ---- found by the in-session adversarial pass, after the cold panels shipped -
+
+
+def test_a_fifo_at_the_registry_path_cannot_hang_the_render(tmp_path):
+    """A NAMED PIPE IS NOT A REGULAR FILE, AND open() BLOCKS ON IT.
+
+    This is the 2026-09-22 lesson recurring in a new file: a size cap that
+    bounds the READ does not help, because the hang is in `open()` — a FIFO
+    with no writer blocks there forever, before a single byte is bounded.
+    `read_benches` runs on the board page's render path, so one FIFO would hang
+    that request and, with enough hits, the threadpool behind every route.
+
+    The guard is a REGULAR-FILE check before the open, which is what marks.py
+    already does (`stat.S_ISREG`). Defeating change: reverting to `path.open()`
+    guarded only by a byte cap — which is what this unit shipped first, while
+    its docstring claimed the cap closed exactly this hole.
+    """
+    os.mkfifo(tmp_path / BENCHES_FILE)
+    benches, err = read_benches(tmp_path)      # must RETURN, not block
+    assert benches == [] and err
+
+
+def test_a_directory_at_the_registry_path_is_an_error_not_a_crash(tmp_path):
+    (tmp_path / BENCHES_FILE).mkdir()
+    benches, err = read_benches(tmp_path)
+    assert benches == [] and err
+
+
+@pytest.mark.parametrize("encoded", ["%00", "%0a", "%0d", "%09", "%1b"])
+def test_a_control_character_is_not_an_addressable_booth(encoded):
+    """`unquote` happily produces a NUL or a newline, and neither can name a
+    real directory. Left unfiltered they reach `is_dir()` (which raises
+    ValueError on an embedded NUL on some paths), the refusal message the CLI
+    prints, and the marker the board renders. Defeating change: dropping the
+    control-character clause — the `%2e%2e` and `%2f` rows above stay green
+    under it, so this needs its own."""
+    assert booth_target(f"http://h:8090/b/{encoded}/") is None
+
+
+def test_normalization_is_idempotent(tmp_path):
+    """LOAD-BEARING for `bench state <id|url>` and `bench rm <id|url>`: both
+    normalize whatever they are handed, so an id must normalize to itself or
+    addressing a bench by the id the registry stores would miss it. Defeating
+    change: any rule that rewrites an already-normalized form."""
+    for u in (GITEA_EIGHT + TALK_FIVE + [
+        "http://x.test/", "http://x.test:8080/p/", "https://x.test/?a=1",
+        "HTTP://X.Test:80/Some%20Path/?b=2&a=1#frag",
+    ]):
+        once = normalize_bench_url(u)
+        assert normalize_bench_url(once) == once, u
+
+
+def test_a_failed_write_leaves_no_scratch_file(tmp_path, monkeypatch):
+    """The temp file is named per-pid so two writers cannot share it, but a
+    write that dies between create and replace would strand it beside the
+    registry forever. Defeating change: dropping the cleanup."""
+    import booth.benches as B
+    upsert_bench(tmp_path, "http://a.test/", "one", "o")
+    real = B.os.replace
+
+    def boom(src, dst):
+        raise OSError("disk full")
+    monkeypatch.setattr(B.os, "replace", boom)
+    with pytest.raises(OSError):
+        upsert_bench(tmp_path, "http://b.test/", "two", "o")
+    monkeypatch.setattr(B.os, "replace", real)
+    strays = [p.name for p in tmp_path.iterdir() if ".tmp" in p.name]
+    assert not strays, strays
+    # and the prior registry is intact — a failed write destroys nothing
+    assert [b.name for b in read_benches(tmp_path)[0]] == ["one"]
+
+
+def test_an_ipv6_literal_keeps_its_brackets():
+    """`urlsplit().hostname` strips them, and a netloc rebuilt from it is not
+    another spelling of the URL — it is a broken one, so a re-post never
+    matches the row the operator means to update. Defeating change: rebuilding
+    netloc from `hostname` with no re-wrap, which is what this shipped as."""
+    assert normalize_bench_url("http://[::1]:8080/a") == "http://[::1]:8080/a"
+    assert normalize_bench_url("http://[2001:DB8::1]/p") == "http://[2001:db8::1]/p"
+    assert normalize_bench_url("HTTP://[::1]:80/p") == "http://[::1]/p"
+    # An UNBRACKETED IPv6 netloc is refused with a reason, not repaired:
+    # `urlsplit(...).port` raises on `::1:8080` because it cannot tell the
+    # address from the port — which is precisely why the brackets exist. The
+    # refusal is the honest answer; guessing where the address ends would be
+    # inventing an identity out of an ambiguous string.
+    with pytest.raises(ValueError):
+        normalize_bench_url("http://::1:8080/a")
+
+
+def test_deeply_nested_json_does_not_escape_the_read(tmp_path):
+    """RecursionError is neither ValueError nor OSError, so it went straight
+    past `read_benches`'s except pair and 500'd the page the function exists to
+    protect. The byte cap does not help: 200k open brackets is 200 KB, well
+    inside it. Defeating change: dropping the RecursionError arm."""
+    (tmp_path / BENCHES_FILE).write_text("[" * 200_000)
+    benches, err = read_benches(tmp_path)
+    assert benches == [] and err
+
+
+def test_the_panel_renders_on_an_EMPTY_board(tmp_path):
+    """The panel is gated on PAGE IDENTITY, not page content. Gating on
+    `board or benches` hid the panel and its registration form exactly when the
+    board was empty and the registry absent — the state a new deployment starts
+    in, and the one where "no benches registered yet" is most worth saying.
+    Defeating change: any content-derived gate."""
+    _board(tmp_path, "")
+    body = _client(tmp_path).get("/b/links/").text
+    assert "no benches registered yet" in body
+    assert "bench-add" in body, "the registration form vanished with the panel"
+
+
+def test_the_panel_shows_when_a_bench_was_added(tmp_path):
+    """INV-N/What renders: the contract says the panel shows the date it was
+    added; `b.added` appeared nowhere in the template and no test asked. All
+    four cold arms found this independently."""
+    _board(tmp_path, ROW_REF)
+    upsert_bench(tmp_path, "https://talk.test/", "talk", "o")
+    added = read_benches(tmp_path)[0][0].added[:10]
+    assert added in _client(tmp_path).get("/b/links/").text
+
+
+@pytest.mark.parametrize("url,expected", BOOTH_URL_TABLE)
+def test_the_dead_marker_classifies_the_SAME_table(tmp_path, url, expected):
+    """INV-2 names RENDER-LEVEL agreement, and the marker tests never ran the
+    table — three hand-written rows with no query between them, so a marker
+    that stopped calling `booth_target` and treated `?q=1` as "not a booth"
+    stayed green while disagreeing with the CLI. Caught by the cold panel.
+
+    Every row whose target does not exist on disk must be marked dead; every
+    non-booth row must not be."""
+    _board(tmp_path, f"- [r]({url}) <sub>· x · 2026-09-01 00:00</sub>\n")
+    marked = bool(_dead_rows(_client(tmp_path).get("/b/links/").text))
+    assert marked is (expected is not None), (url, expected)
+
+
+def test_updated_is_replaced_and_added_is_not(tmp_path, monkeypatch):
+    """The other half of `test_added_survives_reregistration_updated_does_not`,
+    which asserted only the half in the first clause of its own name.
+
+    The stamp has SECOND resolution, so a fast test cannot tell a replaced
+    `updated` from a frozen one by comparing real clocks — `>=` passes either
+    way, which is a falsifier that cannot fail. The clock is driven instead, so
+    "was it rewritten" is answerable. Cold panel, hulda F12.
+
+    Defeating change: carrying `updated` forward from the prior record the way
+    `added` is carried, which every real-clock assertion in this file survives.
+    """
+    import booth.benches as B
+    ticks = iter(["2026-01-01T00:00:00+00:00",
+                  "2026-06-06T06:06:06+00:00",
+                  "2026-12-31T23:59:59+00:00"])
+    monkeypatch.setattr(B, "_now", lambda: next(ticks))
+
+    first, _ = upsert_bench(tmp_path, "http://a.test/", "one", "o")
+    assert first.added == first.updated == "2026-01-01T00:00:00+00:00"
+
+    second, _ = upsert_bench(tmp_path, "http://a.test/", "two", "o")
+    assert second.added == "2026-01-01T00:00:00+00:00", "added must survive an upsert"
+    assert second.updated == "2026-06-06T06:06:06+00:00", "updated must be replaced"
+
+    # A STATE CHANGE IS A MUTATION and bumps it too — this is what the contract
+    # was amended to say, after the panel read "most recent upsert" literally.
+    third = set_bench_state(tmp_path, normalize_bench_url("http://a.test/"), "retired")
+    assert third.added == "2026-01-01T00:00:00+00:00"
+    assert third.updated == "2026-12-31T23:59:59+00:00"
+
+
+def test_a_registration_cannot_make_the_registry_unreadable(tmp_path):
+    """Cold contract panel, hulda solo: the write path permitted a file the
+    reader then refuses on size — so the LAST bench somebody added would be the
+    one that made every other bench invisible, and the write that did it
+    reported success.
+
+    Defeating change: dropping the size check from `_write_all`. The reader is
+    lenient about damage and deliberately NOT lenient about size; a writer
+    ignoring a limit its own reader enforces manufactures exactly the state
+    that leniency exists to survive."""
+    from booth.benches import BENCHES_MAX_BYTES, NAME_MAX
+    n = 0
+    while True:
+        n += 1
+        try:
+            upsert_bench(tmp_path, f"http://h{n}.test/{'p' * 1800}", "x" * NAME_MAX, "o")
+        except ValueError as exc:
+            assert "past" in str(exc) and str(BENCHES_MAX_BYTES) in str(exc)
+            break
+        assert n < 500, "never hit the cap; widen the fixture"
+    # THE REGISTRY IS STILL READABLE, and still holds everything that fit.
+    benches, err = read_benches(tmp_path)
+    assert err is None, err
+    assert len(benches) == n - 1
+
+
+def test_an_over_long_stored_url_is_damage_not_a_silent_clip(tmp_path):
+    """Cold contract panel, 4-of-4 on cap semantics: "applied at the read" did
+    not say TRUNCATE or REFUSE, and the code had picked truncate for every
+    field. For `name` and `owner` that is right — they are display budgets and
+    clipping costs a few characters in a panel row. For `url` it is wrong:
+    INV-7 promises the click goes to the posted address byte for byte, and a
+    clipped URL keeps that promise in the type system while breaking it in the
+    browser. Defeating change: routing `url` back through `_cap`."""
+    from booth.benches import URL_MAX
+    long_url = "http://a/" + "p" * (URL_MAX + 10)
+    _write_raw(tmp_path, json.dumps({"http://a/": {
+        "url": long_url, "name": "n", "owner": "o", "state": "live"}}))
+    benches, err = read_benches(tmp_path)
+    assert err and benches == [], "an over-long url was clipped into a dead anchor"
+
+
+def test_name_and_owner_ARE_clipped_at_the_read(tmp_path):
+    """The other half of the same rule, so the asymmetry is pinned in both
+    directions rather than asserted in one."""
+    from booth.benches import NAME_MAX, OWNER_MAX
+    _write_raw(tmp_path, json.dumps({"http://a/": {
+        "url": "http://a/", "name": "n" * 500, "owner": "o" * 500, "state": "live"}}))
+    benches, err = read_benches(tmp_path)
+    assert err is None
+    assert len(benches[0].name) == NAME_MAX and len(benches[0].owner) == OWNER_MAX
+
+
+def test_the_benches_panel_is_not_nested_inside_a_span(tmp_path):
+    """Cold bug-hunt panel, 3-of-4, seat-confirmed by byte offset in the live
+    document: the panel `<div>` had landed INSIDE the booth header's
+    `<span class="sub">`, because the insertion matched the first
+    `{% if board %}` in the template rather than the block-level one.
+
+    A `<div>` inside a `<span>` is invalid HTML — the parser closes the span
+    implicitly and hoists the div out, orphaning the rest of the sub-line. It
+    renders "fine" in the sense that nothing 500s, which is exactly why no
+    other test in this file could see it.
+
+    Checked the way the seat checked it: by offset. Defeating change: moving
+    the panel back above the sub-span's close."""
+    _board(tmp_path, ROW_REF)
+    upsert_bench(tmp_path, "https://talk.test/", "talk", "o")
+    body = _client(tmp_path).get("/b/links/").text
+    sub_open = body.index('<span class="sub">')
+    sub_close = body.index("</span>", body.index("· ", sub_open))
+    panel = body.index('<div class="benches">')
+    assert not (sub_open < panel < sub_close), (
+        f"the benches div (offset {panel}) sits inside the sub span "
+        f"({sub_open}..{sub_close})")
+
+
+def test_a_symlinked_booth_is_dead_to_the_marker_as_it_is_to_the_page(tmp_path):
+    """Cold bug-hunt panel, 3-of-4: `_booth_exists` used a bare `is_dir()`
+    while `resolve_booth` resolves and requires the parent to BE the data root.
+    They disagreed on a symlink — the marker called a booth pointing outside
+    the root alive while the page 404s it, so the row rendered healthy and the
+    link was dead. The worst of both, and invisible.
+
+    Defeating change: dropping the containment check from `_booth_exists`."""
+    outside = tmp_path.parent / f"outside-{tmp_path.name}"
+    outside.mkdir()
+    try:
+        (tmp_path / "escapee").symlink_to(outside, target_is_directory=True)
+    except OSError:
+        pytest.skip("no symlink support here")
+    _board(tmp_path, "- [x](http://h:8090/b/escapee/) <sub>· a · 2026-09-01 00:00</sub>\n")
+    c = _client(tmp_path)
+    body = c.get("/b/links/")
+    assert body.status_code == 200
+    # the page's own verdict on that name, which the marker must agree with
+    assert c.get("/b/escapee/").status_code == 404
+    assert _dead_rows(body.text), "the marker called a booth alive that the page 404s"

tests/test_booth.py

@@ -16,7 +16,7 @@ from booth.app import (
     remove_link_entry,
     toggle_pin,
     booth_age_seconds,
-    FAVICON_LINK,
+    EMBED_SCRIPT_TAG,
     KEEP_MARKER,
     build_gallery,
     classify,
@@ -30,7 +30,6 @@ from booth.app import (
     render_doc,
     safe_upload_name,
     sweep_once,
-    wrap_verbatim_html,
 )
 
 PICKUP_RE = re.compile(r"^(\d{1,2}-[a-z]+|[a-z]+-\d{1,2})$")
@@ -459,58 +458,21 @@ def test_view_nonviewable_redirects_to_raw(client):
     assert r.headers["location"] == "/b/run1/data.bin"
 
 
-# ---- verbatim-index.html wrapper --------------------------------------------
+# ---- verbatim-index.html serving -------------------------------------------
+#
+# U3 replaced the injection wrapper with a declared seam. The five `test_wrap_*`
+# tests and `test_verbatim_booth_wrapped_with_back_chip` that stood here tested
+# `wrap_verbatim_html` — six regexes hunting a head-ish seam for a favicon and a
+# body-ish seam for a chip, plus the doctype and charset-window constraints they
+# threaded. None of those constraints can be violated by an append, so there is
+# nothing left of them to assert. What replaced them lives in tests/test_embed.py
+# (the payload, the one appended tag, whole-body equality for a declaring page)
+# and tests/test_embed_browser.py (the mount, in a real DOM).
+#
+# What stays here is what did NOT change: the file route is still raw.
 
 
-def test_wrap_injects_chip_and_favicon():
+def test_verbatim_booth_is_served_with_the_seam(client):
-    html = "<html><head><title>Brief</title></head><body><h1>REPORT</h1></body></html>"
-    out = wrap_verbatim_html(html)
-    assert 'class="booth-nav-home"' in out       # floating back chip
-    assert 'href="/"' in out                      # points at the main booth index
-    assert "all booths" in out
-    assert FAVICON_LINK in out                    # favicon inherited
-    assert "<h1>REPORT</h1>" in out               # original content preserved
-    # favicon lands in the head, chip lands in the body
-    assert out.index(FAVICON_LINK) < out.index("</head>")
-    assert out.index("booth-nav-home") > out.index("<body>")
-
-
-def test_wrap_respects_existing_favicon():
-    html = '<html><head><link rel="icon" href="data:image/png;base64,AAAA"></head><body>x</body></html>'
-    out = wrap_verbatim_html(html)
-    assert FAVICON_LINK not in out                # the page's own icon wins
-    assert out.count('rel="icon"') == 1
-    assert 'class="booth-nav-home"' in out        # chip is still added
-
-
-def test_wrap_bare_fragment_appends_chip():
-    out = wrap_verbatim_html("<h1>bare fragment</h1>")  # no doctype/head/body
-    assert 'class="booth-nav-home"' in out
-    assert out.rstrip().endswith("</style>")            # chip appended at the end
-    assert FAVICON_LINK in out                           # no doctype -> safe to prepend the icon
-    assert out.index(FAVICON_LINK) < out.index("bare")   # icon ahead of content (implied head)
-
-
-def test_wrap_no_head_injects_favicon():
-    out = wrap_verbatim_html("<body><h1>no head</h1></body>")
-    assert 'class="booth-nav-home"' in out
-    assert FAVICON_LINK in out                           # injected even without an explicit <head>
-
-
-def test_wrap_compact_doctype_stays_first():
-    # the real-booth shape: compact HTML, no explicit head/body. The injection must
-    # not push anything ahead of the doctype (quirks mode) or past the charset window.
-    html = "<!doctype html><meta charset=utf-8><title>T</title><style>body{margin:0}</style><h1>REPORT</h1>"
-    out = wrap_verbatim_html(html)
-    assert out.lstrip().lower().startswith("<!doctype")       # doctype still first -> standards mode
-    assert FAVICON_LINK in out
-    assert out.index(FAVICON_LINK) < out.index("<h1>")         # icon in the implied head, before content
-    assert out.index("charset") < 1024                         # charset meta stays in the detection window
-    assert 'class="booth-nav-home"' in out
-    assert out.index("booth-nav-home") > out.index("<h1>REPORT</h1>")  # chip appended after content
-
-
-def test_verbatim_booth_wrapped_with_back_chip(client):
     c, data = client
     d = data / "brief"
     d.mkdir()
@@ -518,13 +480,11 @@ def test_verbatim_booth_wrapped_with_back_chip(client):
     r = c.get("/b/brief/")
     assert r.status_code == 200
     assert "BRIEF" in r.text                       # content preserved
-    assert 'class="booth-nav-home"' in r.text      # back chip injected
+    assert r.text.endswith(EMBED_SCRIPT_TAG)       # ...and the seam, appended
-    assert 'href="/"' in r.text
-    assert 'rel="icon"' in r.text                  # favicon inherited
 
 
 def test_verbatim_index_raw_file_route_unwrapped(client):
-    # the file route (/b/<name>/index.html) still serves the raw bytes — the chip
+    # the file route (/b/<name>/index.html) still serves the raw bytes — the seam
     # only rides on the booth view (/b/<name>/), so downloads/assets stay verbatim
     c, data = client
     d = data / "brief"
@@ -532,7 +492,7 @@ def test_verbatim_index_raw_file_route_unwrapped(client):
     (d / "index.html").write_text("<html><body><h1>BRIEF</h1></body></html>")
     r = c.get("/b/brief/index.html")
     assert r.status_code == 200
-    assert "booth-nav-home" not in r.text
+    assert "_booth/embed.js" not in r.text
 
 
 # ---- .md / .txt in-booth doc viewer -----------------------------------------
@@ -781,6 +741,10 @@ def test_releasing_a_board_RESETS_its_ttl_clock(tmp_path):
 
     (kept / KEEP_MARKER).unlink()
 
+    # Unlinking the sentinel by hand, which is what this test is about: the
+    # directory-entry change is what moves the clock. Releasing through the
+    # ROUTE now also records a view, so the behaviour is stated rather than
+    # incidental — `test_releasing_a_board_RECORDS_A_VIEW` in test_lifetime.py.
     assert booth_age_seconds(kept) < 60, "unlink bumped the dir mtime"
     assert sweep_once(tmp_path, ttl_seconds=3600) == [], "so it is NOT swept yet"
     assert kept.exists()
@@ -1537,7 +1501,13 @@ def test_booth_page_offers_keep_when_ephemeral_and_release_when_kept(client):
     _png(d / "x.png")
 
     body = c.get("/b/bo/").text
-    assert "☆ keep" in body and "release" not in body.split("boothhead")[1][:900]
+    # Sliced on the ELEMENT, not the bare word: `boothhead` has appeared in the
+    # stylesheet this page carries since long before this assertion, so
+    # `split("boothhead")[1]` was reading CSS and passing on luck. It went red
+    # the first time a new rule landed above the old one (U5's .prov), which is
+    # the only reason anybody noticed. Same assertion, aimed at the markup.
+    head = body.split('class="boothhead"')[1][:900]
+    assert "☆ keep" in body and "release" not in head
 
     c.post("/b/bo/keep", data={"next": "/b/bo/"}, follow_redirects=False)
     body = c.get("/b/bo/").text

tests/test_cli.py

@@ -0,0 +1,661 @@
+"""`scripts/booth` — the surface every fleet session actually calls.
+
+It had no tests at all, which the 2026-09-22 bug-hunt panel found the hard way:
+its guard-strength table returned UNVERIFIED for every CLI claim because nothing
+in the suite executes the script. Two of that round's findings live in here.
+
+These run the real script under the real system `python3` with no venv, which
+also makes them a live check on INV-1 (stdlib-only): a third-party import in
+`marks.py` fails here the same way it fails on a fleet host.
+"""
+import json
+import os
+import pathlib
+import subprocess
+
+import pytest
+
+SCRIPT = pathlib.Path(__file__).parent.parent / "scripts" / "booth"
+
+# Exit codes the verbs promise. 0 is a successful read; a reader that CRASHED
+# must never be one of the meaningful codes, or a caller cannot tell "no" from
+# "broken" — which is the whole finding.
+OK, UNANSWERED, NO_SUCH_PICK, READER_FAILED = 0, 1, 2, 3
+
+
+def run(data, *args, **kw):
+    env = {**os.environ, "BOOTH_DATA_DIR": str(data), "BOOTH_URL": "http://booth.invalid"}
+    return subprocess.run([str(SCRIPT), *args], capture_output=True, text=True,
+                          env=env, timeout=30, **kw)
+
+
+@pytest.fixture
+def booth(tmp_path):
+    b = tmp_path / "b"
+    b.mkdir()
+    return tmp_path, b
+
+
+def _declare(booth_dir, mark_id="winner"):
+    import sys
+    sys.path.insert(0, str(pathlib.Path(__file__).parent.parent))
+    from booth.marks import declare_pick
+    declare_pick(booth_dir, mark_id,
+                 {"prompt": "Which one?", "options": ["A", "B"]})
+
+
+def test_marks_prints_one_json_document(booth):
+    """`booth marks <name>` is a read. Its stdout is parsed by the session that
+    called it, so it has to be ONE document — and exit 0, because the read
+    succeeded. Whether a pick is open is in the payload's `open` list, which is
+    where a caller should read it from."""
+    data, b = booth
+    _declare(b)
+    r = run(data, "marks", "b")
+    assert r.returncode == OK, r.stderr
+    doc = json.loads(r.stdout)
+    assert doc["open"] == ["winner"]
+
+
+def test_marks_wait_prints_once_not_once_per_poll(booth):
+    """`--wait` polls every 2 s and printed the whole document on every pass, so
+    a capture held several concatenated JSON values and `jq` could not read any
+    of them. The wait is a wait; the print is the result."""
+    data, b = booth
+    _declare(b)
+    import sys
+    sys.path.insert(0, str(pathlib.Path(__file__).parent.parent))
+    from booth.marks import answer_pick
+
+    # Answer it after the first poll so --wait genuinely loops at least once.
+    r = subprocess.Popen([str(SCRIPT), "marks", "b", "--wait", "20"],
+                         stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True,
+                         env={**os.environ, "BOOTH_DATA_DIR": str(data),
+                              "BOOTH_URL": "http://booth.invalid"})
+    import time
+    time.sleep(3)
+    answer_pick(b, "winner", "A")
+    out, err = r.communicate(timeout=30)
+    assert r.returncode == OK, err
+    json.loads(out)          # ONE document, or this raises
+
+
+def test_marks_reports_a_reader_failure_instead_of_printing_garbage(booth):
+    """A traceback on stdout with exit 0 is the worst of both: the caller's `jq`
+    sees success and gets nothing. A read that could not happen is its own
+    answer and gets its own code."""
+    data, b = booth
+    (b / ".marks.json").write_bytes(b"\xff\xfe not utf-8 at all")
+    r = run(data, "marks", "b")
+    assert r.returncode == READER_FAILED, f"rc={r.returncode} out={r.stdout!r}"
+
+
+def test_answer_distinguishes_a_crash_from_an_unanswered_pick(booth):
+    """`answer` funnelled a reader crash and "not yet answered" through the same
+    exit 1, so `--wait` spun for the full hour on a broken file and then blamed
+    the operator for not answering."""
+    data, b = booth
+    _declare(b)
+    r = run(data, "answer", "b", "winner")
+    assert r.returncode == UNANSWERED
+
+    (b / ".marks.json").write_bytes(b"\xff\xfe not utf-8 at all")
+    r = run(data, "answer", "b", "winner", "--wait", "6")
+    assert r.returncode == READER_FAILED, (
+        "a crash was read as 'unanswered' and waited out the timeout"
+    )
+
+
+def test_answer_on_a_note_id_says_no_such_pick(booth):
+    """`answer` matched on id alone while the web route filters on shape, so a
+    note id was reported 'unanswered' and polled forever — a question that could
+    never be answered because it was never a question."""
+    data, b = booth
+    import sys
+    sys.path.insert(0, str(pathlib.Path(__file__).parent.parent))
+    from booth.marks import write_note
+    write_note(b, "a.png", "just a note")
+
+    r = run(data, "answer", "b", "note-1")
+    assert r.returncode == NO_SUCH_PICK
+    assert "no such pick" in r.stderr
+
+
+# ---- U5: self-announcing booths ---------------------------------------------
+
+
+def _manifest(booth_dir):
+    import sys
+    sys.path.insert(0, str(pathlib.Path(__file__).parent.parent))
+    from booth.manifest import read_manifest
+    return read_manifest(booth_dir)
+
+
+def test_new_announces_the_booth(tmp_path):
+    """`$ALTHING_HANDLE` is the whole provenance story: the session already has
+    it, so the booth can say who made it without anybody typing a name."""
+    env = {**os.environ, "ALTHING_HANDLE": "shutter-dev"}
+    r = subprocess.run([str(SCRIPT), "new", "r18-ab", "--why", "pick the winner"],
+                       capture_output=True, text=True, timeout=30,
+                       env={**env, "BOOTH_DATA_DIR": str(tmp_path),
+                            "BOOTH_URL": "http://booth.invalid"})
+    assert r.returncode == 0, r.stderr
+    m = _manifest(tmp_path / "r18-ab")
+    assert m.handle == "shutter-dev"
+    assert m.why == "pick the winner"
+
+
+def test_new_without_a_why_is_still_legal(tmp_path):
+    """The flags are optional and existing call sites keep working. A booth
+    that says only who made it is still a booth that said something."""
+    r = subprocess.run([str(SCRIPT), "new", "scratch"], capture_output=True,
+                       text=True, timeout=30,
+                       env={**os.environ, "ALTHING_HANDLE": "booth-dev",
+                            "BOOTH_DATA_DIR": str(tmp_path),
+                            "BOOTH_URL": "http://booth.invalid"})
+    assert r.returncode == 0, r.stderr
+    m = _manifest(tmp_path / "scratch")
+    assert m.handle == "booth-dev" and m.why == ""
+
+
+def test_add_announces_and_still_copies_the_files(tmp_path):
+    """`add` is the verb most sessions actually use — it creates the booth AND
+    fills it — so the why has to ride on it or it rides nowhere."""
+    src = tmp_path / "src"
+    src.mkdir()
+    (src / "a.txt").write_text("content")
+    r = subprocess.run([str(SCRIPT), "add", "r18-ab", str(src / "a.txt"),
+                        "--why", "second pass", "--title", "R18 A/B"],
+                       capture_output=True, text=True, timeout=30,
+                       env={**os.environ, "ALTHING_HANDLE": "booth-dev",
+                            "BOOTH_DATA_DIR": str(tmp_path),
+                            "BOOTH_URL": "http://booth.invalid"})
+    assert r.returncode == 0, r.stderr
+    assert (tmp_path / "r18-ab" / "a.txt").read_text() == "content"
+    m = _manifest(tmp_path / "r18-ab")
+    assert m.why == "second pass" and m.title == "R18 A/B"
+
+
+def test_add_re_announcing_keeps_the_original_created(tmp_path):
+    """The common shape: `new` opens the booth, `add` drops the second batch and
+    sharpens the why. The booth appeared once."""
+    env = {**os.environ, "ALTHING_HANDLE": "booth-dev",
+           "BOOTH_DATA_DIR": str(tmp_path), "BOOTH_URL": "http://booth.invalid"}
+    src = tmp_path / "a.txt"
+    src.write_text("x")
+    subprocess.run([str(SCRIPT), "new", "b", "--why", "first"], check=True,
+                   capture_output=True, timeout=30, env=env)
+    first = _manifest(tmp_path / "b").created
+    subprocess.run([str(SCRIPT), "add", "b", str(src), "--why", "sharper"],
+                   check=True, capture_output=True, timeout=30, env=env)
+
+    after = _manifest(tmp_path / "b")
+    assert after.created == first
+    assert after.why == "sharper"
+
+
+def test_the_link_board_announces_itself_as_the_booths_own(tmp_path):
+    """No exemption list. The standing board is made by the service and posted
+    to by seventeen handles, so no single agent owns it — `booth` is the
+    truthful answer, and it keeps the rule to one line."""
+    r = subprocess.run([str(SCRIPT), "link", "http://example.invalid", "a thing"],
+                       capture_output=True, text=True, timeout=30,
+                       env={**os.environ, "ALTHING_HANDLE": "booth-dev",
+                            "BOOTH_DATA_DIR": str(tmp_path),
+                            "BOOTH_URL": "http://booth.invalid"})
+    assert r.returncode == 0, r.stderr
+    m = _manifest(tmp_path / "links")
+    assert m is not None and m.handle == "booth"
+    assert m.why
+
+
+def test_the_flags_can_sit_on_either_side_of_the_files(tmp_path):
+    """`booth add b *.png --why "..."` and `booth add b --why "..." *.png` both
+    work. A glob is usually last and a flag usually after it, but nothing
+    enforces that and a session should not have to remember which."""
+    src = tmp_path / "a.png"
+    src.write_bytes(b"x")
+    env = {**os.environ, "ALTHING_HANDLE": "booth-dev",
+           "BOOTH_DATA_DIR": str(tmp_path), "BOOTH_URL": "http://booth.invalid"}
+    for name, args in (("after", ["add", "after", str(src), "--why", "w"]),
+                       ("before", ["add", "before", "--why", "w", str(src)])):
+        r = subprocess.run([str(SCRIPT), *args], capture_output=True, text=True,
+                           timeout=30, env=env)
+        assert r.returncode == 0, r.stderr
+        assert _manifest(tmp_path / name).why == "w"
+        assert (tmp_path / name / "a.png").exists(), "the files stopped being copied"
+
+
+def test_a_why_survives_quotes_and_non_ascii_and_is_flattened(tmp_path):
+    """The reason this goes through manifest.py instead of printf-ing JSON from
+    the shell: a why containing a quote, a backslash or a newline is not an edge
+    case, it is a sentence somebody wrote. Newlines flatten because the field
+    renders inside a card's sub-line."""
+    r = subprocess.run(
+        [str(SCRIPT), "new", "b", "--why", 'he said "pick v3" — line1\nline2 · ünï'],
+        capture_output=True, text=True, timeout=30,
+        env={**os.environ, "ALTHING_HANDLE": "booth-dev",
+             "BOOTH_DATA_DIR": str(tmp_path), "BOOTH_URL": "http://booth.invalid"})
+    assert r.returncode == 0, r.stderr
+    why = _manifest(tmp_path / "b").why
+    assert why == 'he said "pick v3" — line1 line2 · ünï'
+
+
+def test_a_flag_with_no_value_does_not_eat_the_booth_name(tmp_path):
+    """`booth new b --why` with nothing after it must not consume `b` as the
+    value and then create a booth called nothing. Usage, and no directory."""
+    r = subprocess.run([str(SCRIPT), "new", "b", "--why"], capture_output=True,
+                       text=True, timeout=30,
+                       env={**os.environ, "BOOTH_DATA_DIR": str(tmp_path),
+                            "BOOTH_URL": "http://booth.invalid"})
+    assert r.returncode == 2
+    assert "usage:" in r.stderr
+    assert not (tmp_path / "b").exists()
+
+
+def test_a_bare_add_does_not_wipe_the_why_the_new_set(tmp_path):
+    """`booth new x --why "..."` then `booth add x out/*.png` is THE sequence,
+    and the second call must not erase the first one's sentence. The module
+    distinguishes omitted from empty; the shell has to carry that distinction
+    across, which means an UNSET variable, not an empty one."""
+    env = {**os.environ, "ALTHING_HANDLE": "booth-dev",
+           "BOOTH_DATA_DIR": str(tmp_path), "BOOTH_URL": "http://booth.invalid"}
+    src = tmp_path / "a.png"
+    src.write_bytes(b"x")
+
+    subprocess.run([str(SCRIPT), "new", "b", "--why", "pick the denoiser",
+                    "--title", "R18 A/B"],
+                   check=True, capture_output=True, timeout=30, env=env)
+    subprocess.run([str(SCRIPT), "add", "b", str(src)],
+                   check=True, capture_output=True, timeout=30, env=env)
+
+    m = _manifest(tmp_path / "b")
+    assert m.why == "pick the denoiser", "a bare `booth add` wiped the why"
+    assert m.title == "R18 A/B"
+
+
+def test_an_explicitly_empty_why_still_clears_it(tmp_path):
+    """Omitted means unchanged; supplied-and-empty means the poster meant to
+    take it back. Both have to be reachable from the shell."""
+    env = {**os.environ, "ALTHING_HANDLE": "booth-dev",
+           "BOOTH_DATA_DIR": str(tmp_path), "BOOTH_URL": "http://booth.invalid"}
+    subprocess.run([str(SCRIPT), "new", "b", "--why", "wrong"], check=True,
+                   capture_output=True, timeout=30, env=env)
+    subprocess.run([str(SCRIPT), "new", "b", "--why", ""], check=True,
+                   capture_output=True, timeout=30, env=env)
+
+    assert _manifest(tmp_path / "b").why == ""
+
+
+def test_answer_and_marks_agree_about_what_open_means(tmp_path):
+    """U2 made `_is_open` THE openness predicate — "nothing else may spell this
+    out" — and `booth answer`'s reader spelled it out anyway, as
+    `if m.answer is None`. So a PARTIALLY answered pick read as done to
+    `answer` and still-open to `marks --wait`: one verb returns the half-filled
+    form and the other blocks on the same booth at the same instant.
+
+    Found 2/4. The two verbs are the session's whole view of the loop, and a
+    session that asks both gets two answers.
+    """
+    import sys
+    sys.path.insert(0, str(pathlib.Path(__file__).parent.parent))
+    from booth.marks import answer_pick, declare_pick
+
+    b = tmp_path / "b"
+    b.mkdir()
+    declare_pick(b, "batch", {
+        "title": "R18",
+        "questions": [
+            {"key": "q1", "prompt": "One?", "options": ["keep", "drop"]},
+            {"key": "q2", "prompt": "Two?", "options": ["keep", "drop"]},
+        ],
+    })
+    answer_pick(b, "batch", {"q1": "keep", "q2": None})      # partial
+
+    env = {**os.environ, "BOOTH_DATA_DIR": str(tmp_path),
+           "BOOTH_URL": "http://booth.invalid"}
+    marks = subprocess.run([str(SCRIPT), "marks", "b"], capture_output=True,
+                           text=True, timeout=30, env=env)
+    answer = subprocess.run([str(SCRIPT), "answer", "b", "batch"],
+                            capture_output=True, text=True, timeout=30, env=env)
+
+    still_open = "batch" in json.loads(marks.stdout)["open"]
+    assert still_open, "a partial answer stopped counting as open"
+    assert answer.returncode == UNANSWERED, (
+        "`answer` called a partially-answered pick done while `marks` called it open"
+    )
+
+
+def test_answer_does_not_poll_forever_on_a_pick_that_cannot_be_answered(tmp_path):
+    """The mirror failure. A pick whose declaration went bad hydrates with
+    `error` set, which makes it NOT open — so `marks --wait` returns at once
+    while `answer --wait` polled the full hour against a form the web route
+    refuses with a 400. Nothing was ever going to land."""
+    b = tmp_path / "b"
+    b.mkdir()
+    (b / ".marks.json").write_text(json.dumps({
+        "version": 1,
+        "marks": [{"id": "broken", "shape": "pick", "declaration": {},
+                   "error": "pick has no declaration",
+                   "created": "2026-09-21T00:00:00.000000+00:00"}],
+    }))
+
+    r = subprocess.run([str(SCRIPT), "answer", "b", "broken", "--wait", "8"],
+                       capture_output=True, text=True, timeout=40,
+                       env={**os.environ, "BOOTH_DATA_DIR": str(tmp_path),
+                            "BOOTH_URL": "http://booth.invalid"})
+    assert r.returncode != 0
+    assert "broken" in r.stderr.lower() or "cannot" in r.stderr.lower()
+
+
+# ---- U6: benches ------------------------------------------------------------
+#
+# The CLI half of the unit. `docs/contracts/u6_benches.contract.md`.
+
+REFUSED = 2
+
+# Shared with tests/test_benches.py::BOOTH_URL_TABLE — INV-2 says ONE predicate
+# decides what a booth URL is, and these are the rows the CLI must agree on.
+# A second `/b/` check inlined in the shell for speed goes red HERE.
+from test_benches import BOOTH_URL_TABLE  # noqa: E402
+from booth.benches import normalize_bench_url as normalize_bench_url_cli  # noqa: E402
+
+
+@pytest.mark.parametrize("url,is_booth", [(u, e is not None) for u, e in BOOTH_URL_TABLE])
+def test_link_refuses_exactly_what_booth_target_matches(booth, url, is_booth):
+    """INV-2. Defeating change: a `case "$url" in *':8090/b/'*)` in the shell,
+    which would classify the host-agnostic and percent-encoded rows differently
+    from the Python predicate the board's dead marker uses."""
+    data, _ = booth
+    r = run(data, "link", url, "a description")
+    assert (r.returncode == REFUSED) is is_booth, (url, r.returncode, r.stderr)
+
+
+def test_a_refused_link_writes_nothing_at_all(booth):
+    """INV-3. Defeating change: putting the refusal AFTER the `mkdir -p` /
+    announce block, which is where it would naturally land if written without
+    thinking. Asserting only that links.md lacks the row would PASS under that
+    change — so this asserts the board directory does not exist."""
+    data, _ = booth
+    board = data / "links"
+    assert not board.exists()
+    before = sorted(p.name for p in data.iterdir())
+    r = run(data, "link", "http://10.100.10.50:8090/b/some-booth/", "nope")
+    assert r.returncode == REFUSED
+    assert not board.exists(), "a refused link created the board directory"
+    # NOTHING AT ALL, not just no board. Asserting only `links/`'s absence let
+    # a refusal that touched `.benches.lock` (or any other sidecar) on its way
+    # out stay green — the cold panel's vacuity pass named exactly that.
+    assert sorted(p.name for p in data.iterdir()) == before, "a refused link wrote something"
+
+
+def test_the_refusal_names_the_alternative(booth):
+    """The teaching moment belongs at the point of use: 17 handles have the
+    muscle memory, and a bare 'refused' sends them to a human."""
+    data, _ = booth
+    r = run(data, "link", "http://10.100.10.50:8090/b/some-booth/", "nope")
+    out = r.stderr + r.stdout
+    assert "--why" in out and "some-booth" in out
+
+
+def test_a_reference_bookmark_is_still_a_link(booth):
+    """The board keeps its residual job. Measured: ~14 of the 35 distinct
+    non-booth targets are repos, model cards and docs, for which the board is
+    the right and only home. A second refusal would break that."""
+    data, _ = booth
+    r = run(data, "link", "https://gitea.phasefinal.com/vh/peedlar", "the repo")
+    assert r.returncode == OK, r.stderr
+    assert "the repo" in (data / "links" / "links.md").read_text()
+
+
+def test_bench_add_is_an_upsert(booth):
+    data, _ = booth
+    for i in range(3):
+        r = run(data, "bench", "add", "https://talk.nh3.phasefinal.com:8092/", f"talk v{i}")
+        assert r.returncode == OK, r.stderr
+    r = run(data, "bench", "ls")
+    assert r.returncode == OK, r.stderr
+    assert r.stdout.count("talk v") == 1 and "talk v2" in r.stdout
+    # THE ID, WHOLE AND UNTRUNCATED, because it is the locator `bench state`
+    # and `bench rm` take. An earlier version of this test had a docstring
+    # claiming `bench ls` prints ids and asserted nothing of the kind, while
+    # the code printed a url truncated to 52 columns — a claim standing in for
+    # evidence, which is how the drift would have survived CI. Found by all
+    # four cold arms independently.
+    bid = normalize_bench_url_cli("https://talk.nh3.phasefinal.com:8092/")
+    assert bid in r.stdout, r.stdout
+    # and what ls prints is addressable, end to end
+    line = [l for l in r.stdout.splitlines() if "talk v2" in l][0]
+    printed_id = line.split()[-1]
+    assert run(data, "bench", "state", printed_id, "promoted").returncode == OK
+
+
+def test_bench_verbs_round_trip(booth):
+    data, _ = booth
+    assert run(data, "bench", "add", "http://x.test/", "ex").returncode == OK
+    assert run(data, "bench", "state", "http://x.test/", "promoted").returncode == OK
+    assert "promoted" in run(data, "bench", "ls").stdout
+    assert run(data, "bench", "rm", "http://x.test/").returncode == OK
+    assert "ex" not in run(data, "bench", "ls").stdout
+
+
+def test_bench_state_and_rm_take_an_id_or_a_url(booth):
+    """`bench ls` prints ids; the operator has the URL. BOTH must address.
+
+    This used to invoke both verbs with the URL only, twice, while its docstring
+    claimed it covered the id — the same claim-not-evidence shape as the `ls`
+    docstring. A raw URL whose normalization DIFFERS from it is used, so the two
+    columns are genuinely distinct inputs. Cold panel, regin F8.
+    """
+    data, _ = booth
+    raw = "HTTP://X.Test:80/p/?b=2&a=1#frag"
+    bid = normalize_bench_url_cli(raw)
+    assert bid != raw, "pick a URL whose normalization actually differs"
+    run(data, "bench", "add", raw, "ex")
+    # by the ID the registry stores
+    assert run(data, "bench", "state", bid, "retired").returncode == OK
+    assert "retired" in run(data, "bench", "ls").stdout
+    # and by the RAW URL the operator has in their scrollback
+    assert run(data, "bench", "state", raw, "live").returncode == OK
+    assert "live" in run(data, "bench", "ls").stdout
+    assert run(data, "bench", "rm", raw).returncode == OK
+    run(data, "bench", "add", raw, "ex again")
+    assert run(data, "bench", "rm", bid).returncode == OK
+    assert "ex" not in run(data, "bench", "ls").stdout
+
+
+def test_bench_add_refuses_a_bad_url_with_the_reason(booth):
+    data, _ = booth
+    r = run(data, "bench", "add", "ftp://x.test/f", "ex")
+    assert r.returncode != OK
+    assert "http" in (r.stderr + r.stdout).lower()
+
+
+def test_bare_bench_names_the_bench_verbs(booth):
+    """Seam review SR-6: `bench` is the first two-word verb in this script, and
+    falling through to the generic usage hides which word was wrong."""
+    data, _ = booth
+    r = run(data, "bench")
+    assert r.returncode != OK
+    assert "add" in r.stderr and "import" in r.stderr
+
+
+def _seed_board(data):
+    board = data / "links"
+    board.mkdir(parents=True, exist_ok=True)
+    (board / "links.md").write_text(
+        "- [a booth](http://10.100.10.50:8090/b/gone/) <sub>· x · 2026-09-01 00:00</sub>\n"
+        "- [talk](https://talk.nh3.phasefinal.com:8092/) <sub>· x · 2026-09-01 00:00</sub>\n"
+        "- [talk again](https://talk.nh3.phasefinal.com:8092/) <sub>· x · 2026-09-02 00:00</sub>\n"
+        "- [a repo](https://gitea.phasefinal.com/vh/peedlar) <sub>· x · 2026-09-03 00:00</sub>\n"
+        "- [bad](ftp://x.test/f) <sub>· x · 2026-09-04 00:00</sub>\n"
+    )
+    return board
+
+
+def test_import_writes_nothing_without_apply(booth):
+    """INV-8. A proposal that writes is not a proposal."""
+    data, _ = booth
+    board = _seed_board(data)
+    before = (board / "links.md").read_text()
+    r = run(data, "bench", "import")
+    assert r.returncode == OK, r.stderr
+    assert not (data / ".benches.json").exists()
+    assert (board / "links.md").read_text() == before
+
+
+def test_import_classifies_into_three_groups(booth):
+    data, _ = booth
+    _seed_board(data)
+    out = run(data, "bench", "import").stdout
+    assert "gone" in out          # the booth row, skipped
+    assert "talk" in out          # a candidate
+    assert "ftp://x.test/f" in out  # refused, with its reason
+    # THE RAW URL BESIDE THE NORMALIZED ID, which is the entire point of the
+    # proposal: five rows of `talk` collapsing to one is only checkable if you
+    # can see which raw URLs produced the one id. This printed the description
+    # instead, so the collapse was invisible in the one place it had to be
+    # visible. All four cold arms found it.
+    assert out.count("https://talk.nh3.phasefinal.com:8092/") >= 2, out
+
+
+def test_bare_apply_refuses_and_writes_nothing(booth):
+    """THE SELECTION GAP — all four cold contract-review arms, independently.
+
+    `--apply` used to register every candidate, while the same contract says
+    roughly 14 of 35 are reference bookmarks that must STAY on the board. That
+    made the write path do the exact thing the unit's own rationale calls
+    impossible — tell a bench from a bookmark by its URL — silently, to rows
+    that belong where they are. The dry-run prints ids; `--apply` takes the
+    ones the operator names, and refuses without them.
+
+    Defeating change: restoring the register-everything branch."""
+    data, _ = booth
+    _seed_board(data)
+    r = run(data, "bench", "import", "--apply")
+    assert r.returncode == REFUSED
+    assert "needs the ids" in r.stderr
+    assert not (data / ".benches.json").exists(), "a bare --apply wrote the registry"
+
+
+def test_apply_refuses_an_id_that_is_not_a_candidate(booth):
+    data, _ = booth
+    _seed_board(data)
+    r = run(data, "bench", "import", "--apply", "http://not-on-the-board/")
+    assert r.returncode == REFUSED
+    assert "not a candidate id" in r.stderr
+    assert not (data / ".benches.json").exists()
+
+
+def test_apply_registers_ONLY_the_named_ids(booth):
+    """The bookmark stays a bookmark unless the operator says otherwise."""
+    data, _ = booth
+    _seed_board(data)
+    talk = normalize_bench_url_cli("https://talk.nh3.phasefinal.com:8092/")
+    assert run(data, "bench", "import", "--apply", talk).returncode == OK
+    ls = run(data, "bench", "ls").stdout
+    assert "peedlar" not in ls, "an unnamed candidate was registered anyway"
+    assert len([l for l in ls.splitlines() if "talk" in l]) == 1
+
+
+def test_import_apply_collapses_the_repost(booth):
+    data, _ = booth
+    _seed_board(data)
+    talk = normalize_bench_url_cli("https://talk.nh3.phasefinal.com:8092/")
+    repo = normalize_bench_url_cli("https://gitea.phasefinal.com/vh/peedlar")
+    assert run(data, "bench", "import", "--apply", talk, repo).returncode == OK
+    ls = run(data, "bench", "ls").stdout
+    # ONE ROW, counted by line: "talk" appears in both the name and the
+    # hostname, so a substring count would read 2 for a correctly collapsed row.
+    assert len([l for l in ls.splitlines() if "talk" in l]) == 1, ls
+    assert "peedlar" in ls
+    assert "gone" not in ls, "a booth row was imported as a bench"
+
+
+def test_nothing_in_the_unit_touches_links_md(booth):
+    """INV-8. Defeating change: `import --apply` tidying up the rows it
+    consumed. The whole CLI surface runs against one board and the file must
+    come out byte-identical."""
+    import hashlib
+    data, _ = booth
+    board = _seed_board(data)
+    before = hashlib.sha256((board / "links.md").read_bytes()).hexdigest()
+    run(data, "link", "http://10.100.10.50:8090/b/x/", "refused")
+    run(data, "bench", "import")
+    run(data, "bench", "import", "--apply")          # refused, writes nothing
+    run(data, "bench", "import", "--apply",
+        normalize_bench_url_cli("https://talk.nh3.phasefinal.com:8092/"))
+    run(data, "bench", "add", "http://new.test/", "new")
+    run(data, "bench", "state", "http://new.test/", "retired")
+    run(data, "bench", "ls")          # a read verb can truncate too
+    run(data, "bench", "rm", "http://new.test/")
+    after = hashlib.sha256((board / "links.md").read_bytes()).hexdigest()
+    assert before == after
+
+
+def test_link_fails_CLOSED_when_the_booth_check_cannot_run(booth, tmp_path):
+    """A guard that fails open is not a guard. If `booth.links` cannot be
+    imported, `booth link` must post NOTHING and say why — not append the row
+    it could not classify, and not abort with a bare traceback.
+
+    Defeating change: dropping the `|| pred_rc=$?` handling, which under
+    `set -e` aborts with a Python traceback (safe, but unactionable), or
+    treating a failed check as "not a booth" (unsafe — fails open)."""
+    data, _ = booth
+    lone = tmp_path / "lone" / "scripts"
+    lone.mkdir(parents=True)
+    (lone / "booth").write_text(SCRIPT.read_text())
+    (lone / "booth").chmod(0o755)
+    r = subprocess.run([str(lone / "booth"), "link", "https://ok.test/x", "a bookmark"],
+                       capture_output=True, text=True, cwd="/tmp", timeout=30,
+                       env={**os.environ, "BOOTH_DATA_DIR": str(data),
+                            "BOOTH_URL": "http://booth.invalid"})
+    assert r.returncode != OK
+    assert "could not check" in r.stderr, r.stderr
+    assert not (data / "links" / "links.md").exists(), "a row landed despite an unusable check"
+
+
+def test_a_credential_never_reaches_the_board(booth):
+    """`normalize_bench_url` refuses userinfo for a bench; `booth link` was the
+    door this unit did not touch, and the board renders on an unauthenticated
+    LAN surface. Cold contract panel, groa solo. A deliberate small widening of
+    the unit, named rather than smuggled."""
+    data, _ = booth
+    r = run(data, "link", "https://user:hunter2@x.test/p", "leaky")
+    assert r.returncode != OK
+    assert "credentials" in r.stderr
+    assert not (data / "links").exists(), "a credentialed URL created the board"
+
+
+def test_the_append_happens_INSIDE_the_lock(booth):
+    """Cold bug-hunt panel, hulda solo. `flock LOCK printf ... >> board` reads
+    as locked and is not: the SHELL opens the append fd while parsing, before
+    flock acquires. A concurrent `unlink` rewriting the board in that window
+    replaces the inode, the old fd keeps pointing at the unlinked one, and the
+    append succeeds, reports success, and vanishes.
+
+    Proved by holding the lock: if the open were outside it, `booth link` would
+    write and exit while blocked. Defeating change: reverting to the bare
+    `flock LOCK printf ... >>` form, under which this test writes the row.
+    """
+    import fcntl
+    data, _ = booth
+    board = data / "links"
+    board.mkdir(parents=True)
+    (board / "links.md").write_text("")
+    lock = board / ".links.lock"
+    lock.touch()
+    with lock.open("r+") as lf:
+        fcntl.flock(lf, fcntl.LOCK_EX)
+        try:
+            # subprocess.run directly: `run()` pins timeout=30 itself.
+            with pytest.raises(subprocess.TimeoutExpired):
+                subprocess.run(
+                    [str(SCRIPT), "link", "https://ok.test/x", "blocked"],
+                    capture_output=True, text=True, timeout=5,
+                    env={**os.environ, "BOOTH_DATA_DIR": str(data),
+                         "BOOTH_URL": "http://booth.invalid"})
+        finally:
+            fcntl.flock(lf, fcntl.LOCK_UN)
+    assert (board / "links.md").read_text() == "", \
+        "the row was appended while another writer held the lock"

tests/test_embed.py

@@ -0,0 +1,527 @@
+"""U3 — the declared embed seam, server side.
+
+The Booth used to reach into a verbatim report with ten regular expressions: six
+to find somewhere to hang a favicon and a chip, four to substitute rendered ask
+markup into the author's own tags. This unit replaces all of it with a seam the
+page declares:
+
+    <script src="/_booth/embed.js" defer></script>
+
+What is tested here is the SERVER half — the payload that crosses the seam, the
+one static asset, and the single conditional append that is now the only thing
+the Booth does to author HTML. The half that mounts fragments into a live DOM
+lives in tests/test_embed_browser.py, because no amount of string assertion can
+see whether a form actually submits.
+
+Contract: docs/contracts/u3_declared_embed_seam.contract.md
+"""
+import json
+import os
+import time
+
+import pytest
+from fastapi.testclient import TestClient
+
+from booth.app import EMBED_SCRIPT_TAG, EMBED_SRC, create_app
+from booth.marks import answer_pick, declare_pick, set_flag, write_note
+
+DECLARED = f'<!doctype html><title>r</title><body>hi<script src="{EMBED_SRC}" defer></script></body>'
+
+
+@pytest.fixture
+def client(tmp_path):
+    app = create_app(tmp_path, ttl_hours=24, start_sweeper=False)
+    return TestClient(app), tmp_path
+
+
+def _pick(booth, stem="winner", **kw):
+    doc = {"prompt": "Which render wins?", "options": ["A — baseline", "B — async"]}
+    doc.update(kw)
+    booth.mkdir(parents=True, exist_ok=True)
+    declare_pick(booth, stem, doc)
+    return booth
+
+
+def _multi(booth, stem="batch"):
+    booth.mkdir(parents=True, exist_ok=True)
+    declare_pick(booth, stem, {"title": "Round one", "questions": [
+        {"key": "r1", "prompt": "First?", "options": ["a", "b"]},
+        {"key": "r2", "prompt": "Second?", "options": ["a", "b"]},
+    ]})
+    return booth
+
+
+# ---- slice 1: the payload ----------------------------------------------------
+
+
+def test_embed_payload_carries_a_fragment_for_every_shape(client):
+    c, data = client
+    _multi(data / "b")
+    body = c.get("/b/b/embed.json").json()
+    assert body["booth"] == "b"
+    assert "home" not in body, "a value nothing reads is a second copy waiting to drift"
+    assert body["favicon"].startswith("data:image/svg+xml,")
+    (m,) = body["marks"]
+    assert m["id"] == "batch" and m["error"] is None
+    assert "First?" in m["whole"] and "Second?" in m["whole"]
+    assert 'action="/b/b/answer"' in m["submit"]
+    assert [q["key"] for q in m["questions"]] == ["r1", "r2"]   # declaration order
+    assert "First?" in m["questions"][0]["html"]
+    assert 'type="radio"' in m["questions"][0]["html"]
+
+
+def test_a_single_question_pick_has_one_question_with_a_null_key(client):
+    """SR-2. `normalize_ask` gives a single-question ask `key: None`, so the
+    payload cannot key questions by name — JSON would write that as "null" and
+    invent a name. Every one-question ask in the fleet hits this."""
+    c, data = client
+    _pick(data / "b")
+    (m,) = c.get("/b/b/embed.json").json()["marks"]
+    assert [q["key"] for q in m["questions"]] == [None]
+    assert "Which render wins?" in m["questions"][0]["html"]
+
+
+def test_marks_are_ordered_by_creation_before_id(client):
+    """`created` LEADS. The fixture makes creation order and alphabetical order
+    disagree, because a fixture where they agree cannot tell the stated rule
+    from a plain id sort — which is what the first version of this test did, and
+    what a panel caught by reading the fixture rather than the assertion."""
+    c, data = client
+    b = _pick(data / "b", "zebra")
+    _pick(b, "alpha")
+    raw = json.loads((b / ".marks.json").read_text())
+    stamps = {"zebra": "2026-09-22T10:00:00.000000-07:00",   # first
+              "alpha": "2026-09-22T11:00:00.000000-07:00"}   # second
+    for e in raw["marks"]:
+        e["created"] = stamps[e["id"]]
+    (b / ".marks.json").write_text(json.dumps(raw))
+    assert [m["id"] for m in c.get("/b/b/embed.json").json()["marks"]] == ["zebra", "alpha"]
+
+
+def test_the_id_is_only_the_tie_break(client):
+    """And with `created` equal, the id decides — so two marks written in the
+    same second cannot swap between renders."""
+    c, data = client
+    b = _pick(data / "b", "zebra")
+    _pick(b, "alpha")
+    raw = json.loads((b / ".marks.json").read_text())
+    for e in raw["marks"]:
+        e["created"] = "2026-09-22T10:00:00.000000-07:00"
+    (b / ".marks.json").write_text(json.dumps(raw))
+    assert [m["id"] for m in c.get("/b/b/embed.json").json()["marks"]] == ["alpha", "zebra"]
+
+
+def test_open_is_computed_by_the_server_not_the_page(client):
+    """INV-4. The chip count follows `open_marks`, which is the ONE openness
+    predicate — a half-answered multi-question pick is still open."""
+    c, data = client
+    b = _multi(data / "b")
+    assert c.get("/b/b/embed.json").json()["open"] == ["batch"]
+    answer_pick(b, "batch", {"r1": "a"})
+    assert c.get("/b/b/embed.json").json()["open"] == ["batch"]   # partial is OPEN
+    answer_pick(b, "batch", {"r1": "a", "r2": "b"})
+    assert c.get("/b/b/embed.json").json()["open"] == []
+
+
+def test_the_payload_carries_picks_only(client):
+    """SR-4. Notes and flags never reach it, which is also what keeps a flag's
+    `flag:<target>` id — the one mark id containing the anchor separator — out
+    of a payload whose specs split on the first colon."""
+    c, data = client
+    b = _pick(data / "b")
+    (b / "shot.png").write_bytes(b"x")
+    write_note(b, None, "a remark")
+    set_flag(b, "shot.png", True)
+    assert [m["id"] for m in c.get("/b/b/embed.json").json()["marks"]] == ["winner"]
+
+
+def test_a_damaged_marks_file_does_not_500_the_report(client):
+    c, data = client
+    b = _pick(data / "b")
+    (b / ".marks.json").write_text("{not json")
+    r = c.get("/b/b/embed.json")
+    assert r.status_code == 200
+    assert r.json()["marks"] == [] and r.json()["error"]
+
+
+def test_a_broken_pick_offers_whole_and_nothing_else(client):
+    c, data = client
+    b = _pick(data / "b")
+    raw = json.loads((b / ".marks.json").read_text())
+    raw["marks"][0]["declaration"] = {"prompt": "p"}       # no options -> AskError
+    (b / ".marks.json").write_text(json.dumps(raw))
+    (m,) = c.get("/b/b/embed.json").json()["marks"]
+    assert m["error"] and m["questions"] == [] and m["submit"] == ""
+    assert "broken ask" in m["whole"]
+
+
+def test_the_payload_does_not_record_a_view(client):
+    """`booth_view` already recorded the look, above both of its early returns.
+    A script's fetch of the page it is already on must not count a second time
+    or reset the TTL on machinery instead of on the operator."""
+    c, data = client
+    b = _pick(data / "b")
+    (b / "index.html").write_text(DECLARED)
+    c.get("/b/b/")
+    before = os.stat(b / ".viewed").st_mtime_ns
+    time.sleep(0.01)
+    c.get("/b/b/embed.json")
+    assert os.stat(b / ".viewed").st_mtime_ns == before
+
+
+def test_embed_payload_404s_for_an_unknown_booth(client):
+    c, _ = client
+    assert c.get("/b/nope/embed.json").status_code == 404
+
+
+# ---- slice 2: the one static asset -------------------------------------------
+
+
+def test_embed_js_is_served_as_javascript(client):
+    c, _ = client
+    r = c.get(EMBED_SRC)
+    assert r.status_code == 200
+    assert r.headers["content-type"].startswith("text/javascript")
+    assert "data-booth-mark" in r.text
+
+
+def test_embed_js_does_not_hot_reload_from_disk(tmp_path):
+    """INV-5, and the 2026-09-21 lesson restated. A live asset editable under a
+    running process is how 19 of 25 booths hit 500 with the Python from 22:03
+    and the templates from 23:40. One rule in this repo: nothing takes effect
+    until you restart."""
+    import pathlib
+
+    import booth.app as app_mod
+
+    src = pathlib.Path(app_mod.__file__).parent / "static" / "embed.js"
+    original = src.read_text()
+    app = create_app(tmp_path, ttl_hours=24, start_sweeper=False)
+    c = TestClient(app)
+    try:
+        # Poisoned BEFORE the first request, not between two of them. The
+        # earlier shape passed for a route that read the file lazily and cached
+        # on first use — which is not "read once at startup", and is exactly the
+        # staleness this invariant exists to forbid.
+        src.write_text("/* POISONED */\n")
+        first = c.get(EMBED_SRC).text
+        assert "POISONED" not in first, "embed.js is read at request time, not at startup"
+        assert first == original
+        assert c.get(EMBED_SRC).text == first
+    finally:
+        src.write_text(original)
+
+
+# ---- slice 3: what the Booth does to author HTML -----------------------------
+
+
+def test_declaring_page_is_served_untouched(client):
+    """INV-1. Whole-body equality, not a substring absence: the promise is that
+    NOTHING is added, and an absence assertion cannot tell a clean page from one
+    carrying something nobody thought to look for."""
+    c, data = client
+    b = _pick(data / "b")
+    (b / "index.html").write_text(DECLARED)
+    assert c.get("/b/b/").text == DECLARED
+
+
+def test_undeclared_page_gains_only_the_tag(client):
+    """INV-2. Appended, so the source is a strict prefix — nothing is inserted,
+    nothing is prepended, and neither the doctype nor the charset window moves."""
+    c, data = client
+    b = _pick(data / "b")
+    src = "<!doctype html><meta charset=utf-8><title>r</title><h1>REPORT</h1>"
+    (b / "index.html").write_text(src)
+    out = c.get("/b/b/").text
+    # The literal, not just the constant: `out == src + EMBED_SCRIPT_TAG` also
+    # holds when EMBED_SCRIPT_TAG is the empty string, which is a mutation this
+    # test exists to catch. A panel found it by reading the assertion, not the
+    # code.
+    assert out == src + '<script src="/_booth/embed.js" defer></script>'
+    assert out == src + EMBED_SCRIPT_TAG
+    assert len(out) > len(src)
+    assert out.lower().lstrip().startswith("<!doctype")
+    assert out.index("charset") < 1024
+
+
+def test_a_booth_with_no_marks_still_gets_the_seam(client):
+    """The seam carries the way home and the icon too, so it is not conditional
+    on there being an ask — the old chip was not either."""
+    c, data = client
+    (data / "b").mkdir()
+    (data / "b" / "index.html").write_text("<h1>bare fragment</h1>")
+    assert c.get("/b/b/").text == "<h1>bare fragment</h1>" + EMBED_SCRIPT_TAG
+
+
+def test_a_declaring_page_is_served_BYTE_for_byte(client):
+    """INV-1, at the level the promise is actually made.
+
+    The first version read the file with `read_text()`, which opens in
+    universal-newline mode: a report written with CRLF came back with LF, and
+    `errors="replace"` turned any non-UTF-8 byte into U+FFFD. A declaring page
+    was NOT served as its author wrote it — the headline promise — and the
+    original test could not see it, because its fixture was LF-only ASCII.
+    Found by a cross-frontier bug-hunt panel.
+    """
+    c, data = client
+    b = _pick(data / "b")
+    src = (b'<!doctype html>\r\n<title>r</title>\r\n<body>caf\xe9 \xff\r\n'
+           b'<script src="/_booth/embed.js" defer></script>\r\n</body>')
+    (b / "index.html").write_bytes(src)
+    r = c.get("/b/b/")
+    assert r.content == src, "the operator's document was edited on the way out"
+    assert b"\r\n" in r.content and b"\xff" in r.content
+
+
+def test_an_undeclared_page_keeps_every_byte_and_gains_the_tag(client):
+    """INV-2, same level: the source is a BYTE-exact prefix of the response."""
+    c, data = client
+    b = _pick(data / "b")
+    src = b'<!doctype html>\r\n<title>r</title>\r\n<body>caf\xe9 \xff\r\n</body>'
+    (b / "index.html").write_bytes(src)
+    r = c.get("/b/b/")
+    assert r.content == src + EMBED_SCRIPT_TAG.encode("utf-8")
+    assert r.content.startswith(src)
+
+
+def test_a_wrongly_shaped_answer_costs_its_pick_not_the_report(client):
+    """`marks_for` hydrates `{"answer": {"answers": []}}` with no error — the
+    JSON is well formed, the SHAPE is not — and the template then asks a list
+    for `.get`. This endpoint renders every pick on every load of the operator's
+    report, so an unguarded raise here is the whole seam gone while `hold_read`
+    calls the file perfectly readable. Verified reachable, not assumed."""
+    c, data = client
+    b = data / "b"
+    b.mkdir(parents=True, exist_ok=True)
+    declare_pick(b, "batch", {"title": "T", "questions": [
+        {"key": "r1", "prompt": "A?", "options": ["x", "y"]},
+        {"key": "r2", "prompt": "B?", "options": ["x", "y"]}]})
+    _pick(b, "healthy")
+    raw = json.loads((b / ".marks.json").read_text())
+    for e in raw["marks"]:
+        if e["id"] == "batch":
+            e["answer"] = {"answers": [], "notes": ""}
+    (b / ".marks.json").write_text(json.dumps(raw))
+
+    (b / "index.html").write_text(DECLARED)
+    r = c.get("/b/b/embed.json")
+    assert r.status_code == 200
+    by = {m["id"]: m for m in r.json()["marks"]}
+    # THE PROMISE, NOT THE LAYER. This used to pin the string
+    # `_safe_fragments` produces ("could not be rendered"), which made the test
+    # an assertion about WHICH guard fired. As of the `_hydrate` answer-shape
+    # check, this input is caught one layer earlier and never reaches
+    # `_pick_fragments` at all — the endpoint's promise is unchanged and the
+    # error is better (it names what is wrong with the stored answer instead of
+    # reporting a render failure), so the assertion moved to the promise.
+    # `_safe_fragments` is still the backstop and is still falsified, by
+    # `test_safe_fragments_still_catches_what_hydration_cannot` below.
+    assert by["batch"]["error"], "a wrong-shaped answer reported no error"
+    assert "broken ask" in by["batch"]["whole"]
+    # and the booth's other pick is untouched — one bad entry costs one entry
+    assert by["healthy"]["error"] is None
+    assert "Which render wins?" in by["healthy"]["whole"]
+    assert c.get("/b/b/").status_code == 200
+
+    # ⚠ THE GALLERY AND MARKS PAGES USED TO 500 ON THIS ENTRY, and that was NOT
+    # U3's doing — measured at 42ea67f, the commit before that unit. CLOSED
+    # 2026-09-22 at the hydration boundary rather than by a third copy of this
+    # guard: see tests/test_marks.py
+    # ::test_a_wrong_shaped_answer_is_an_error_at_hydration_not_a_500 and
+    # persistent-memory.d/2026-09-22-a-wrong-shaped-answer-500s-the-gallery.md
+
+
+def test_safe_fragments_still_catches_what_hydration_cannot(client):
+    """U3's `_safe_fragments` guard, kept falsifiable after `_hydrate` took its
+    natural trigger away.
+
+    The answer-shape check in `_hydrate` now catches every wrong answer shape
+    reachable from a `.marks.json` — probed 2026-09-22: `answers` as a list, a
+    string or null all become hydration errors, and a wrong-typed VALUE inside
+    `answers` renders without raising, because Jinja absorbs attribute access
+    on a non-mapping. **No natural input reaches `_safe_fragments` by this
+    route any more**, and a test that kept pretending one did would assert
+    nothing — which is the failure this suite has now paid for twice.
+
+    So the trigger is synthetic and says so: the shared `_ask_inline` macro
+    module is made to raise. `_pick_fragments` resolves `whole` off that object
+    per call, and `create_app` stashes the environment on `app.state`, so this
+    reaches the very object the closure captured. What it pins is the guard
+    itself — one raising pick costs that pick, never the report.
+
+    Defeating change: removing the try/except in `_safe_fragments`, under which
+    this returns 500.
+    """
+    c, data = client
+    b = data / "b"
+    b.mkdir(parents=True, exist_ok=True)
+    declare_pick(b, "batch", {"prompt": "Which?", "options": ["x", "y"]})
+    (b / "index.html").write_text(DECLARED)
+
+    frag = c.app.state.templates.env.get_template("_ask_inline.html").module
+    real_submit = frag.submit
+
+    def explode(*a, **k):
+        raise RuntimeError("synthetic render failure")
+
+    object.__setattr__(frag, "submit", explode)
+    try:
+        assert frag.submit is explode, "the patch did not take; this test is vacuous"
+        r = c.get("/b/b/embed.json")
+        assert r.status_code == 200, "a raising fragment renderer took the whole report"
+        by = {m["id"]: m for m in r.json()["marks"]}
+        assert by["batch"]["error"] and "could not be rendered" in by["batch"]["error"]
+        assert by["batch"]["whole"], "the fallback rendered nothing at all"
+    finally:
+        object.__setattr__(frag, "submit", real_submit)
+
+    # and the guard is not sticky — with the macro restored, the pick is fine
+    assert c.get("/b/b/embed.json").json()["marks"][0]["error"] is None
+
+
+def test_the_handler_survives_the_failure_it_is_handling(client):
+    """`_safe_fragments` caught a raising `_pick_fragments` and then rebuilt the
+    broken-ask box THROUGH THE SAME MACRO MODULE that had just raised. So when
+    `whole` itself was the broken thing, the handler re-raised and took the
+    whole report — a guard that only worked when the failure was somewhere
+    else.
+
+    Found by accident: the first draft of the falsifier above patched `whole`,
+    and the guard failed rather than caught. Defeating change: removing the
+    inner try/except, under which this returns 500."""
+    c, data = client
+    b = data / "b"
+    b.mkdir(parents=True, exist_ok=True)
+    declare_pick(b, "batch", {"prompt": "Which?", "options": ["x", "y"]})
+    (b / "index.html").write_text(DECLARED)
+
+    frag = c.app.state.templates.env.get_template("_ask_inline.html").module
+    real_whole = frag.whole
+
+    def explode(*a, **k):
+        raise RuntimeError("even the fallback macro is broken")
+
+    object.__setattr__(frag, "whole", explode)
+    try:
+        r = c.get("/b/b/embed.json")
+        assert r.status_code == 200, "the handler re-raised through the broken macro"
+        assert r.json()["marks"][0]["error"]
+    finally:
+        object.__setattr__(frag, "whole", real_whole)
+
+
+def test_no_regex_touches_author_html():
+    """INV-3, and the version that actually falsifies it.
+
+    The first draft of this test name-matched the six deleted patterns. A cold
+    panel pointed out — correctly, and on the contract's CENTRAL promise — that
+    reintroducing the same regex under a new name (`_TAIL_RE`, applied in the
+    verbatim branch) would leave it green. A test that guards names does not
+    guard behaviour, and this repo's own vacuity pass missed it because the
+    mutation it tried was the named one.
+
+    So: `booth/app.py` is allowed EXACTLY ONE regex operation, and it is
+    `ask_form_id`'s `re.sub` over a mark id — not over a page. Any other regex
+    anywhere in the module fails here, whatever it is called. If a future
+    change genuinely needs one, the failure is the conversation: say which
+    string it reads and why it is not author HTML.
+    """
+    import ast
+    import pathlib
+
+    import booth.app as app_mod
+
+    root = pathlib.Path(app_mod.__file__).parent
+    assert not (root / "inline.py").exists(), "booth/inline.py survived U3"
+    tree = ast.parse((root / "app.py").read_text())
+
+    # every node -> the function it sits in, so a finding names its site
+    site = {}
+    for node in ast.walk(tree):
+        if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)):
+            for child in ast.walk(node):
+                site.setdefault(child, node.name)
+
+    # `re` reaches this module ONE way: a plain module-level `import re`. An
+    # alias (`import re as _r`) or a direct name import (`from re import sub`)
+    # would route around the call check below under a name it does not know —
+    # found by re-running the vacuity pass against the FIXED test, which is the
+    # only reason it is here and is the argument for running that pass on a fix
+    # and not only on a draft.
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Import):
+            for a in node.names:
+                assert not (a.name == "re" and a.asname), f"`re` aliased as {a.asname}"
+        elif isinstance(node, ast.ImportFrom):
+            assert node.module != "re", f"names imported from re: {[a.name for a in node.names]}"
+
+    METHODS = {"search", "sub", "subn", "match", "fullmatch", "finditer",
+               "findall", "split", "compile", "escape"}
+    found = []
+    for node in ast.walk(tree):
+        if not (isinstance(node, ast.Call) and isinstance(node.func, ast.Attribute)):
+            continue
+        f = node.func
+        on_re = isinstance(f.value, ast.Name) and f.value.id == "re"
+        on_pattern = (isinstance(f.value, ast.Name) and f.value.id.endswith("_RE")
+                      and f.attr in METHODS)
+        if on_re or on_pattern:
+            found.append((site.get(node, "<module level>"), f.attr))
+
+    assert found == [("ask_form_id", "sub")], (
+        f"booth/app.py performs regex operations outside ask_form_id: {found}"
+    )
+
+    # the six named patterns and the chips are gone, and stay gone
+    assigned = {
+        t.id
+        for node in ast.walk(tree)
+        if isinstance(node, ast.Assign)
+        for t in node.targets
+        if isinstance(t, ast.Name)
+    }
+    gone = {"_ICON_RE", "_HEAD_CLOSE_RE", "_HTML_OPEN_RE", "_DOCTYPE_RE",
+            "_BODY_CLOSE_RE", "_HTML_CLOSE_RE", "_BACK_CHIP", "FAVICON_LINK"}
+    assert not (assigned & gone), f"deleted names are back: {sorted(assigned & gone)}"
+    funcs = {n.name for n in ast.walk(tree) if isinstance(n, ast.FunctionDef)}
+    assert not ({"wrap_verbatim_html", "asks_chip", "inject_asks",
+                 "_insert_before", "_insert_after"} & funcs)
+
+
+def test_a_page_that_only_mentions_the_path_is_not_declaring_it(client):
+    """A report that QUOTES the seam — a code sample, a comment, a sentence
+    about this very feature — is not declaring it, and the Booth's own design
+    reports are the pages most likely to do that. Read as declared, such a page
+    would be served untouched and show no chrome at all, silently.
+
+    The detection therefore fails the other way: an unrecognised spelling gets a
+    duplicate tag, and embed.js mounts once regardless.
+    """
+    c, data = client
+    b = _pick(data / "b")
+    for body in (
+        "<!doctype html><body><p>add <code>/_booth/embed.js</code> to your report</p></body>",
+        "<!doctype html><body><!-- src=/_booth/embed.js --></body>",
+        '<!doctype html><body><script src="/_booth/embed.js?v=2"></script></body>',
+    ):
+        (b / "index.html").write_text(body)
+        assert c.get("/b/b/").text == body + EMBED_SCRIPT_TAG, body
+
+    # and the real declaration, in either quote style, is honoured
+    for decl in (f'<script src="{EMBED_SRC}" defer></script>',
+                 f"<script src='{EMBED_SRC}' defer></script>"):
+        body = f"<!doctype html><body>hi{decl}</body>"
+        (b / "index.html").write_text(body)
+        assert c.get("/b/b/").text == body
+
+
+def test_an_oversize_verbatim_page_is_served_raw(client, monkeypatch):
+    """WRAP_MAX_BYTES survives: a pathological file is still not pulled into
+    memory, and it loses its chrome exactly as it does today."""
+    import booth.app as app_mod
+
+    c, data = client
+    b = _pick(data / "b")
+    (b / "index.html").write_text("<h1>huge</h1>")
+    monkeypatch.setattr(app_mod, "WRAP_MAX_BYTES", 4)
+    assert c.get("/b/b/").text == "<h1>huge</h1>"

tests/test_embed_browser.py

@@ -0,0 +1,539 @@
+"""U3 — the declared embed seam, in a real DOM.
+
+The Python suite can prove what the server OFFERS. It cannot prove where a
+fragment lands, whether the author's own markup survived the mount, or whether
+four radio groups scattered down a report still submit as one POST — and that
+last one is the operator's most important workflow. Before U3 those properties
+were true by construction, because the server did the placing and the `form=`
+bindings were static by the time the page was parsed. Now they are true because
+`/_booth/embed.js` does it in a live document, which is a different kind of
+claim and needs a different kind of test.
+
+So: a real uvicorn on an ephemeral port, a real Chromium.
+
+SKIPS, NEVER FAILS, when playwright or the shared browser is unavailable. The
+box-wide store at /opt/ms-playwright pins specific Chromium revisions and a
+playwright release that wants a newer one dies with an opaque "Executable
+doesn't exist" — see pyproject's version bound. A test layer that goes red for
+an environment reason teaches nothing and trains people to ignore it.
+"""
+import json
+import socket
+import threading
+import time
+
+import pytest
+
+from booth.app import create_app
+
+playwright_api = pytest.importorskip(
+    "playwright.sync_api", reason="playwright is not installed"
+)
+
+
+@pytest.fixture(scope="module")
+def browser():
+    with playwright_api.sync_playwright() as pw:
+        try:
+            b = pw.chromium.launch()
+        except Exception as exc:  # noqa: BLE001 - any launch failure is a skip
+            pytest.skip(f"no usable chromium: {exc}")
+        yield b
+        b.close()
+
+
+@pytest.fixture
+def live(tmp_path):
+    """A real server, because a browser cannot talk to a TestClient."""
+    import uvicorn
+
+    sock = socket.socket()
+    sock.bind(("127.0.0.1", 0))
+    port = sock.getsockname()[1]
+    sock.close()
+
+    app = create_app(tmp_path, ttl_hours=24, start_sweeper=False)
+    config = uvicorn.Config(app, host="127.0.0.1", port=port, log_level="error")
+    server = uvicorn.Server(config)
+    thread = threading.Thread(target=server.run, daemon=True)
+    thread.start()
+    deadline = time.time() + 10
+    while not server.started and time.time() < deadline:
+        time.sleep(0.02)
+    if not server.started:
+        pytest.skip("uvicorn did not come up")
+    try:
+        yield f"http://127.0.0.1:{port}", tmp_path
+    finally:
+        server.should_exit = True
+        thread.join(timeout=10)
+
+
+SEAM = '<script src="/_booth/embed.js" defer></script>'
+
+
+def _multi(booth):
+    from booth.marks import declare_pick
+
+    booth.mkdir(parents=True, exist_ok=True)
+    declare_pick(booth, "batch", {"title": "Round one", "questions": [
+        {"key": "r1", "prompt": "First?", "options": ["keep", "cut"]},
+        {"key": "r2", "prompt": "Second?", "options": ["keep", "cut"]},
+    ]})
+    return booth
+
+
+def _single(booth):
+    from booth.marks import declare_pick
+
+    booth.mkdir(parents=True, exist_ok=True)
+    declare_pick(booth, "winner", {"prompt": "Which render wins?",
+                                   "options": ["A — baseline", "B — async"]})
+    return booth
+
+
+def _open(browser, base, name, html, booth):
+    (booth / "index.html").write_text(html, encoding="utf-8")
+    page = browser.new_page()
+    page.goto(f"{base}/b/{name}/", wait_until="networkidle")
+    return page
+
+
+def _answer(booth):
+    raw = json.loads((booth / ".marks.json").read_text())
+    return raw["marks"][0].get("answer")
+
+
+# ---- the chrome --------------------------------------------------------------
+
+
+def test_a_declaring_page_gets_its_chrome_mounted(browser, live):
+    base, data = live
+    b = _single(data / "b")
+    page = _open(browser, base, "b", f"<!doctype html><title>r</title><body><h1>R</h1>{SEAM}</body>", b)
+    page.wait_for_selector(".booth-nav-home")
+    assert page.locator("h1").inner_text() == "R"            # the report is intact
+    assert page.locator(".booth-nav-home").get_attribute("href").endswith("/")
+    # the favicon question, asked of a parsed document instead of raw text
+    assert page.locator('link[rel="icon"]').count() == 1
+    page.close()
+
+
+def test_a_page_that_never_declared_the_seam_still_mounts(browser, live):
+    """The appended path: every verbatim booth that predates U3 keeps working
+    without its author touching it."""
+    base, data = live
+    b = _single(data / "b")
+    page = _open(browser, base, "b", "<!doctype html><body><h1>OLD</h1></body>", b)
+    page.wait_for_selector(".bk-ask")
+    assert page.locator("h1").inner_text() == "OLD"
+    assert page.locator(".booth-nav-home").count() == 1
+    page.close()
+
+
+def test_a_page_that_declares_its_own_icon_keeps_it(browser, live):
+    base, data = live
+    b = _single(data / "b")
+    page = _open(
+        browser, base, "b",
+        f'<!doctype html><head><link rel="icon" href="data:image/png;base64,AAAA">'
+        f"</head><body>x{SEAM}</body>", b)
+    page.wait_for_selector(".booth-nav-home")
+    icons = page.locator('link[rel="icon"]')
+    assert icons.count() == 1
+    assert icons.get_attribute("href").startswith("data:image/png")
+    page.close()
+
+
+# ---- placement ---------------------------------------------------------------
+
+
+REPORT = f"""<!doctype html><title>audition</title><body>
+<h1>Three voices</h1>
+<section id="lawson"><audio src="a.wav"></audio>
+<div data-booth-ask="batch:r1"></div></section>
+<section id="jo"><audio src="b.wav"></audio>
+<div data-booth-mark="batch:r2"></div></section>
+<div data-booth-ask-submit="batch"></div>
+{SEAM}
+</body>"""
+
+
+def test_each_question_lands_where_the_author_put_it(browser, live):
+    """The 2026-09-09 ruling, enforced in the DOM: the question for a voice sits
+    under that voice, not on another page and not in a pile at the end. Both
+    attribute spellings, because live reports use the older one."""
+    base, data = live
+    b = _multi(data / "b")
+    page = _open(browser, base, "b", REPORT, b)
+    page.wait_for_selector("#lawson .bk-ask")
+    assert page.locator('#lawson input[name="choice.r1"]').count() == 2
+    assert page.locator('#jo input[name="choice.r2"]').count() == 2
+    # nothing spilled to the end of the body: every piece had an anchor
+    assert page.locator("body > .bk-ask").count() == 0
+    assert page.locator("form#bk-ask-form-batch").count() == 1
+    page.close()
+
+
+def test_the_authors_wrapper_and_its_contents_survive_the_mount(browser, live):
+    """The live `dfa-concepts` shape — a non-empty styled wrapper carrying the
+    anchor attribute. The regex this replaced matched the opening tag and
+    SUBSTITUTED it, eating the class and orphaning the heading. beforeend keeps
+    both and puts the radios under the heading, which is what the markup says."""
+    base, data = live
+    b = _multi(data / "b")
+    page = _open(browser, base, "b", (
+        '<!doctype html><body><div class="ask" data-booth-ask="batch:r1">'
+        f"<h3>The one asset that must survive</h3></div>{SEAM}</body>"), b)
+    page.wait_for_selector(".ask .bk-ask")
+    assert page.locator("div.ask").count() == 1                       # class kept
+    assert page.locator(".ask h3").inner_text() == "The one asset that must survive"
+    assert page.locator('.ask input[name="choice.r1"]').count() == 2  # radios inside
+    page.close()
+
+
+def test_an_unplaced_question_is_appended_and_so_is_its_submit(browser, live):
+    """INV-7. A multi-question pick needs EVERY question on submit or the POST is
+    unanswerable: a question that never reaches the page cannot be picked, and
+    a submission with nothing picked at all is refused outright."""
+    base, data = live
+    b = _multi(data / "b")
+    page = _open(browser, base, "b",
+                 f'<!doctype html><body><div data-booth-mark="batch:r1"></div>{SEAM}</body>', b)
+    # attached, not visible: the shared <form> is deliberately empty and so has
+    # no box — the controls that bind to it are what the operator sees.
+    page.wait_for_selector("form#bk-ask-form-batch", state="attached")
+    # BOTH halves. Asserting only the appended r2 let a mutation that silently
+    # swallowed the anchored r1 — while still recording it as placed — pass.
+    assert page.locator('input[name="choice.r1"]').count() == 2   # anchored, mounted
+    assert page.locator('input[name="choice.r2"]').count() == 2   # unplaced, appended
+    assert page.locator("form#bk-ask-form-batch").count() == 1    # submittable
+    page.close()
+
+
+def test_a_page_with_no_anchors_gets_the_whole_ask(browser, live):
+    base, data = live
+    b = _multi(data / "b")
+    page = _open(browser, base, "b", f"<!doctype html><body><p>x</p>{SEAM}</body>", b)
+    page.wait_for_selector(".bk-ask")
+    assert page.locator('input[name="choice.r1"]').count() == 2
+    assert page.locator('input[name="choice.r2"]').count() == 2
+    page.close()
+
+
+def test_an_anchor_naming_no_mark_is_left_alone(browser, live):
+    """A typo'd id stays visible as the author's own empty element rather than
+    being blanked — and the real ask is still never lost."""
+    base, data = live
+    b = _multi(data / "b")
+    page = _open(browser, base, "b",
+                 f'<!doctype html><body><div id="t" data-booth-mark="typo"></div>{SEAM}</body>', b)
+    page.wait_for_selector(".bk-ask")
+    assert page.locator("#t").inner_html().strip() == ""
+    assert page.locator('input[name="choice.r1"]').count() == 2
+    page.close()
+
+
+def test_the_tail_follows_payload_order(browser, live):
+    """INV-6. Creation order and alphabetical order DISAGREE here on purpose:
+    `winner` is created first, `batch` second, so payload order is
+    winner-then-batch while an id sort would give the reverse. The first version
+    of this fixture made the two identical, so sorting the tail alphabetically
+    in JavaScript passed it — caught by a panel reading the fixture."""
+    base, data = live
+    b = _single(data / "b")
+    _multi(b)
+    raw = json.loads((b / ".marks.json").read_text())
+    stamps = {"winner": "2026-09-22T10:00:00.000000-07:00",   # first
+              "batch": "2026-09-22T11:00:00.000000-07:00"}    # second
+    for e in raw["marks"]:
+        e["created"] = stamps[e["id"]]
+    (b / ".marks.json").write_text(json.dumps(raw))
+    page = _open(browser, base, "b", f"<!doctype html><body>{SEAM}</body>", b)
+    page.wait_for_selector(".bk-ask")
+    ids = page.eval_on_selector_all("[id^='bk-ask-']", "els => els.map(e => e.id)")
+    batch = [i for i, v in enumerate(ids) if "batch" in v]
+    winner = [i for i, v in enumerate(ids) if "winner" in v]
+    assert batch and winner, ids
+    # winner (created first) ahead of batch (created second) — the OPPOSITE of
+    # alphabetical, so an id sort cannot pass this.
+    assert max(winner) < min(batch), ids
+    page.close()
+
+
+# ---- the one that actually matters ------------------------------------------
+
+
+def test_a_form_scattered_down_the_report_submits_every_question(browser, live):
+    """THE load-bearing browser test.
+
+    Four radio groups under four different artifacts, one <form> somewhere else
+    entirely, bound only by the HTML5 `form=` attribute — and now inserted into
+    a live document in visual order, which means a control can land before the
+    form it points at. If form-owner resolution does not survive that, the
+    operator fills the whole thing in and the button saves nothing.
+
+    It was true by construction before U3 (static HTML, resolved at parse). It
+    is true by measurement now. That is the trade this test pays for.
+    """
+    base, data = live
+    b = _multi(data / "b")
+    page = _open(browser, base, "b", REPORT, b)
+    page.wait_for_selector("#lawson .bk-ask")
+    page.check('#lawson input[name="choice.r1"][value="keep"]')
+    page.check('#jo input[name="choice.r2"][value="cut"]')
+    with page.expect_navigation():
+        page.click("button.bk-ask-go")
+    ans = _answer(b)
+    assert ans is not None, "the scattered form submitted nothing"
+    assert ans["answers"]["r1"]["choice"] == "keep"
+    assert ans["answers"]["r2"]["choice"] == "cut", \
+        "a question bound by form= did not reach the POST"
+    assert ans["complete"] is True
+    page.close()
+
+
+def test_the_chip_jumps_to_the_first_fragment_of_the_open_ask(browser, live):
+    base, data = live
+    b = _multi(data / "b")
+    page = _open(browser, base, "b", REPORT, b)
+    page.wait_for_selector(".booth-nav-asks")
+    chip = page.locator(".booth-nav-asks")
+    assert chip.inner_text() == "? 1 open ask"
+    target = chip.get_attribute("href")
+    # the EARLIEST match in document order, not merely a match: the report
+    # anchors r1 above r2 above the submit block, so any later one is wrong.
+    ids = page.eval_on_selector_all("[id^='bk-ask-batch']", "els => els.map(e => e.id)")
+    assert ids, "no batch fragment mounted"
+    assert target == "#" + ids[0], (target, ids)
+    page.close()
+
+
+def test_an_anchor_named_like_an_object_property_does_not_kill_the_page(browser, live):
+    """`toString` is a legal mark id (`asks.valid_stem`) and therefore a legal
+    thing for an author to typo into an anchor. Against a plain `{}` lookup it
+    came back as Object.prototype.toString — truthy, so it sailed past the
+    unknown-mark guard and threw on `.questions.length`, aborting placement
+    before the tail and costing the page EVERY ask. One typo, no chrome, no
+    error the operator would see. Found by a cross-frontier panel."""
+    base, data = live
+    b = _multi(data / "b")
+    page = _open(browser, base, "b", (
+        '<!doctype html><body><div id="t" data-booth-mark="toString"></div>'
+        f'<div id="v" data-booth-mark="valueOf:r1"></div>{SEAM}</body>'), b)
+    page.wait_for_selector(".bk-ask")
+    assert page.locator("#t").inner_html().strip() == ""      # left alone
+    assert page.locator("#v").inner_html().strip() == ""      # left alone
+    # and the real ask still mounted, which is what the bug destroyed
+    assert page.locator('input[name="choice.r1"]').count() == 2
+    assert page.locator('input[name="choice.r2"]').count() == 2
+    assert page.locator("form#bk-ask-form-batch").count() == 1
+    page.close()
+
+
+def test_a_question_keyed_like_an_object_property_is_not_swallowed(browser, live):
+    """The mirror of the same bug, on the `placed` set. `constructor` matches
+    `asks._KEY_RE`, and against a plain object an inherited `got.constructor`
+    read as ALREADY PLACED — so a question the author did not anchor was
+    silently dropped from the tail, which is INV-7's whole subject."""
+    from booth.marks import declare_pick
+
+    base, data = live
+    b = data / "b"
+    b.mkdir(parents=True, exist_ok=True)
+    declare_pick(b, "batch", {"title": "Round one", "questions": [
+        {"key": "r1", "prompt": "First?", "options": ["keep", "cut"]},
+        {"key": "constructor", "prompt": "Second?", "options": ["keep", "cut"]},
+    ]})
+    page = _open(browser, base, "b",
+                 f'<!doctype html><body><div data-booth-mark="batch:r1"></div>{SEAM}</body>', b)
+    page.wait_for_selector("form#bk-ask-form-batch", state="attached")
+    assert page.locator('input[name="choice.r1"]').count() == 2
+    assert page.locator('input[name="choice.constructor"]').count() == 2, \
+        "an unplaced question was swallowed by an inherited property"
+    page.close()
+
+
+def test_a_submit_anchor_inside_the_authors_own_form_still_submits(browser, live):
+    """A submit anchor placed inside the author's own `<form>` loses ours: the
+    HTML parser drops a nested form element outright. The controls' `form=`
+    then points at nothing, the button does nothing, and the operator finds out
+    by filling the whole thing in. Found by a cross-frontier bug-hunt panel; the
+    fix is to count the anchor submitted only if the form actually survived, so
+    the tail supplies one at body level where no form encloses it."""
+    base, data = live
+    b = _multi(data / "b")
+    page = _open(browser, base, "b", (
+        '<!doctype html><body>'
+        '<div data-booth-mark="batch:r1"></div>'
+        '<div data-booth-mark="batch:r2"></div>'
+        '<form id="mine" action="/elsewhere">'
+        '<div data-booth-ask-submit="batch"></div></form>'
+        f"{SEAM}</body>"), b)
+    page.wait_for_selector("form#bk-ask-form-batch", state="attached")
+    assert page.locator("form#bk-ask-form-batch").count() == 1
+    page.check('input[name="choice.r1"][value="keep"]')
+    page.check('input[name="choice.r2"][value="cut"]')
+    with page.expect_navigation():
+        page.click("button.bk-ask-go")
+    ans = _answer(b)
+    assert ans is not None, "the button reached no form"
+    assert ans["answers"]["r1"]["choice"] == "keep"
+    assert ans["answers"]["r2"]["choice"] == "cut"
+    page.close()
+
+
+def test_a_broken_pick_shows_its_diagnostic_even_from_a_submit_anchor(browser, live):
+    """A broken pick has no submit block — its `submit` is empty and the
+    diagnostic lives in `whole`. Mounting that empty string and then recording
+    the pick as placed made the tail skip it, so the 'broken ask' box never
+    rendered at the one surface built to show it. A question the session
+    believes it posted has to be visible."""
+    base, data = live
+    b = _single(data / "b")
+    raw = json.loads((b / ".marks.json").read_text())
+    raw["marks"][0]["declaration"] = {"prompt": "p"}      # no options -> AskError
+    (b / ".marks.json").write_text(json.dumps(raw))
+    page = _open(browser, base, "b",
+                 f'<!doctype html><body><div id="s" data-booth-ask-submit="winner">'
+                 f"</div>{SEAM}</body>", b)
+    page.wait_for_selector(".bk-ask")
+    assert page.locator("#s").inner_html().strip() == ""     # anchor left alone
+    # inner_text() is the RENDERED text, and `.bk-ask-tag` is uppercased by CSS —
+    # so assert the diagnostic itself, which is the part that has to reach him.
+    shown = page.locator("body").inner_text()
+    assert "this question could not be read" in shown, shown
+    assert "options" in shown                                # the actual reason
+    page.close()
+
+
+def test_an_author_element_cannot_hijack_the_chip(browser, live):
+    """`<section id="bk-ask-winner-background">` satisfies any id-prefix rule,
+    hyphen boundary included. The chip therefore searches only the elements THIS
+    SCRIPT MOUNTED — the identity the deleted `bk-ask-<id>-top` anchor used to
+    guarantee — and takes the earliest of those in document order."""
+    base, data = live
+    b = _single(data / "b")
+    page = _open(browser, base, "b", (
+        '<!doctype html><body><section id="bk-ask-winner-background">notes</section>'
+        f"<p>report</p>{SEAM}</body>"), b)
+    page.wait_for_selector(".booth-nav-asks")
+    target = page.locator(".booth-nav-asks").get_attribute("href")
+    assert target != "#bk-ask-winner-background"
+    landed = page.locator(target)
+    assert landed.count() == 1
+    assert landed.evaluate("e => e.classList.contains('bk-ask')"), \
+        "the chip jumped to something the Booth did not mount"
+    page.close()
+
+
+def test_the_chip_does_not_jump_to_a_mark_that_merely_shares_a_prefix(browser, live):
+    """A SIBLING MARK's fragment must not take the chip, even when it is earlier
+    in the document. `batch2`'s id starts with `batch`, so the original
+    id-prefix rule could land on it; the mounted-elements rule cannot, because
+    the candidates are partitioned by mark.
+
+    ⚠ The first version of this test put the sibling's fragment AFTER the open
+    mark's, so the right answer was also the first answer and pooling every
+    mark's elements passed it. The vacuity pass caught that; the fixture now
+    puts the sibling FIRST, which is the only arrangement that can tell the two
+    implementations apart.
+    """
+    from booth.marks import declare_pick
+
+    base, data = live
+    b = _multi(data / "b")                       # `batch`, created first
+    declare_pick(b, "batch2", {"prompt": "Unrelated?", "options": ["x", "y"]})
+
+    # batch2 is anchored at the very top; batch is unanchored and so lands in
+    # the tail, at the END of the body. Document order is therefore batch2's
+    # fragments, then batch's.
+    page = _open(browser, base, "b", (
+        '<!doctype html><body><div data-booth-mark="batch2"></div>'
+        f"<p>report</p>{SEAM}</body>"), b)
+    page.wait_for_selector(".booth-nav-asks")
+
+    ids = page.eval_on_selector_all("[id^='bk-ask-']", "els => els.map(e => e.id)")
+    assert any("batch2" in i for i in ids) and any(
+        "batch2" not in i and "batch" in i for i in ids), ids
+    assert ids.index(next(i for i in ids if "batch2" in i)) < \
+        ids.index(next(i for i in ids if "batch2" not in i and "batch" in i)), \
+        f"fixture is wrong: the sibling must come FIRST, got {ids}"
+
+    # `open` is (created, id) -> batch before batch2, so the chip targets batch
+    target = page.locator(".booth-nav-asks").get_attribute("href")
+    assert "batch2" not in target, f"the chip landed on the sibling mark: {target}"
+    assert target.startswith("#bk-ask-batch")
+    assert page.locator(target).count() == 1
+    page.close()
+
+
+def test_the_canonical_attribute_wins_when_both_are_present(browser, live):
+    """`data-booth-mark` is canonical and `data-booth-ask` is the kept alias.
+    An element carrying both is not a case any live report has, but the
+    precedence has to be decided somewhere rather than by selector order."""
+    base, data = live
+    b = _multi(data / "b")
+    page = _open(browser, base, "b", (
+        '<!doctype html><body><div id="a" data-booth-mark="batch:r2" '
+        f'data-booth-ask="batch:r1"></div>{SEAM}</body>'), b)
+    page.wait_for_selector("#a .bk-ask")
+    assert page.locator('#a input[name="choice.r2"]').count() == 2   # canonical
+    assert page.locator('#a input[name="choice.r1"]').count() == 0   # alias ignored
+    # r1 was never placed, so INV-7 still puts it somewhere
+    assert page.locator('input[name="choice.r1"]').count() == 2
+    page.close()
+
+
+def test_the_chip_count_comes_from_the_server(browser, live):
+    """INV-4. A half-answered multi-question pick is STILL OPEN, and the page
+    does not get to have an opinion about that — `open_marks` decides."""
+    from booth.marks import answer_pick
+
+    base, data = live
+    b = _multi(data / "b")
+    answer_pick(b, "batch", {"r1": "keep"})
+    page = _open(browser, base, "b", REPORT, b)
+    page.wait_for_selector(".bk-ask")
+    assert page.locator(".booth-nav-asks").count() == 1
+    answer_pick(b, "batch", {"r1": "keep", "r2": "cut"})
+    page.reload(wait_until="networkidle")
+    page.wait_for_selector(".bk-ask")
+    assert page.locator(".booth-nav-asks").count() == 0
+    page.close()
+
+
+def test_the_chip_follows_a_payload_that_disagrees_with_the_fragments(browser, live):
+    """INV-4, and the version that actually falsifies it.
+
+    The test above uses honest fixtures, so a client that INFERRED openness from
+    the rendered fragments would pass it — the fragments and `open` always agree
+    when the server computes both. A panel pointed out that this never creates
+    the disagreement it claims to test.
+
+    So: intercept the response and make `open` lie. The fragments say fully
+    answered; the payload says two are open. The chip must follow the payload,
+    because the payload is the only thing that decides.
+    """
+    import json as _json
+
+    from booth.marks import answer_pick
+
+    base, data = live
+    b = _multi(data / "b")
+    answer_pick(b, "batch", {"r1": "keep", "r2": "cut"})   # nothing is open
+    (b / "index.html").write_text(f"<!doctype html><body>{SEAM}</body>", encoding="utf-8")
+
+    def lie(route):
+        body = _json.loads(route.fetch().text())
+        assert body["open"] == [], "fixture is not answered; the lie would be true"
+        body["open"] = ["batch", "batch"]
+        route.fulfill(status=200, content_type="application/json",
+                      body=_json.dumps(body))
+
+    page = browser.new_page()
+    page.route("**/embed.json", lie)
+    page.goto(f"{base}/b/b/", wait_until="networkidle")
+    page.wait_for_selector(".booth-nav-asks")
+    assert page.locator(".booth-nav-asks").inner_text() == "? 2 open asks"
+    page.close()

tests/test_items.py

@@ -258,3 +258,61 @@ def test_list_booths_counts_match_the_resolver(tmp_path):
 
     got = list_booths(tmp_path, ttl_seconds=86400)[0]
     assert got["count"] == len(booth_items(b)) == 2
+
+
+# --- U7: the group, derived here and nowhere else -------------------------
+#
+# ⚠ THE RULE IS NOT THE ONE THE CONTRACT FIRST STATED, and the change is
+# measured rather than preferred. The contract's `strip ONE trailing run of
+# digits` yields 24 groups for sindra-bakeoff's 40 images and 27 for sindra's
+# 30 — a rail with one row per tile, which is a second copy of the grid rather
+# than a way through it. Measured against all 17 live booths on 2026-09-22;
+# the numbers are in the contract's rewritten table.
+
+
+def test_group_of_takes_the_first_segment(tmp_path):
+    from booth.items import _group_of
+
+    assert _group_of("00-sheet-c1-market-noon.png") == "00"
+    assert _group_of("m-c1-market-noon-9401.png") == "m"
+    assert _group_of("flag-rear.png") == "flag"
+    assert _group_of("v30-seed8302-HELD.png") == "v30"
+
+
+def test_group_of_destems_only_a_flat_name(tmp_path):
+    """`ac01.png` has no separator, so the digits ARE the separator and the
+    group is `ac`. `v30-seed8302` HAS one, so `v30` survives intact — stripping
+    there would merge v30 with v35, which is the axis that booth is about."""
+    from booth.items import _group_of
+
+    assert _group_of("ac01.png") == "ac"
+    assert _group_of("DSC0001.jpg") == "DSC"
+    assert _group_of("v30-seed8302.png") == "v30"
+    assert _group_of("v35-seed8302.png") == "v35"
+
+
+def test_group_of_is_none_when_there_is_no_prefix(tmp_path):
+    """A stem that is entirely digits has nothing to group on. Inventing one
+    would file every numbered render under the empty string."""
+    from booth.items import _group_of
+
+    assert _group_of("01.png") is None
+    assert _group_of("0042.jpg") is None
+    assert _group_of("-leading.png") is None
+
+
+def test_group_is_derived_from_the_basename_not_the_path(tmp_path):
+    """A booth WITH subdirectories still groups on the filename. Sections and
+    groups are different questions; `Item.section` still carries the path."""
+    from booth.items import _group_of
+
+    assert _group_of("sub/dir/ac01.png") == "ac"
+
+
+def test_booth_items_carries_the_group(tmp_path):
+    b = tmp_path / "g"
+    _touch(b / "ac01.png")
+    _touch(b / "ac02.png")
+    _touch(b / "99.png")
+    got = {it.rel: it.group for it in booth_items(b)}
+    assert got == {"ac01.png": "ac", "ac02.png": "ac", "99.png": None}

tests/test_manifest.py

@@ -0,0 +1,719 @@
+"""U5 — self-announcing booths.
+
+A booth carries `.booth.json` saying who posted it and why, and the index card
+and the booth page render it. Closes job 5 (`Announce`) — the job nobody named,
+whose absence is the measured cause of 145 dead link rows.
+
+See docs/contracts/u5_booth_manifest.contract.md.
+"""
+import ast
+import json
+import os
+import pathlib
+import sys
+
+import pytest
+
+from booth.manifest import MANIFEST_FILE, Manifest, read_manifest, write_manifest
+
+
+# ---- slice 1: the record and its storage ------------------------------------
+
+
+def test_an_announcement_round_trips(tmp_path):
+    b = tmp_path / "r18-ab"
+    b.mkdir()
+    written = write_manifest(b, "booth-dev", why="pick the winning denoiser")
+
+    assert (b / MANIFEST_FILE).is_file()
+    got = read_manifest(b)
+    assert got == written
+    assert got.handle == "booth-dev"
+    assert got.why == "pick the winning denoiser"
+    assert got.error is None
+
+
+def test_the_title_falls_back_to_the_directory_name(tmp_path):
+    """A booth always has a display name. `title` is the one the poster chose
+    when there is one, and the folder name is a perfectly good one when there
+    is not — an empty heading on a card is worse than a plain one."""
+    b = tmp_path / "r18-ab"
+    b.mkdir()
+    assert write_manifest(b, "booth-dev").title == "r18-ab"
+    assert write_manifest(b, "booth-dev", title="R18 A/B").title == "R18 A/B"
+
+
+def test_a_booth_that_never_announced_reads_as_none(tmp_path):
+    """The normal case for every booth that predates this unit, and for every
+    booth that arrives by rsync — the documented path for any host that is not
+    nh3-dev, which never runs the CLI at all."""
+    b = tmp_path / "quiet"
+    b.mkdir()
+    assert read_manifest(b) is None
+    assert read_manifest(tmp_path / "does-not-exist") is None
+
+
+def test_one_line_by_construction_not_by_convention(tmp_path):
+    """`why` renders inside a card's sub-line, so a newline in it would break
+    the card rather than the field. Truncation and newline-stripping happen at
+    the WRITE, so nothing downstream has to remember."""
+    b = tmp_path / "b"
+    b.mkdir()
+    m = write_manifest(b, "booth-dev", why="first line\nsecond line\r\nthird")
+    assert "\n" not in m.why and "\r" not in m.why
+    assert "first line" in m.why and "second line" in m.why
+
+    long = write_manifest(b, "booth-dev", why="x" * 5000)
+    assert len(long.why) <= 200
+    assert len(write_manifest(b, "y" * 500).handle) <= 64
+    assert len(write_manifest(b, "booth-dev", title="t" * 500).title) <= 120
+
+
+# ---- slice 2: the read cannot raise (INV-2) ---------------------------------
+
+
+@pytest.mark.parametrize(
+    "payload",
+    [
+        b"{truncated",                       # not JSON at all
+        b"[]",                               # JSON, wrong shape
+        b'"a string"',                       # JSON, wronger shape
+        b"null",
+        b'{"handle": 7}',                    # right shape, wrong type
+        b'{"why": "no handle here"}',        # the one required field missing
+        b"\xff\xfe not utf-8",
+        b"",
+    ],
+    ids=["truncated", "list", "string", "null", "wrong-type", "no-handle",
+         "not-utf8", "empty"],
+)
+def test_a_damaged_manifest_never_raises(tmp_path, payload):
+    """INV-2. `list_booths` calls this once per booth on every index page load,
+    so a read that can raise is a service-wide outage wearing a single-booth
+    bug's clothes. That is not a hypothetical — a poisoned `.marks.json` did
+    exactly that to `/` and `/healthz` across all 25 booths, and the fix shipped
+    in v0.2.2. The same reader posture, applied before the same mistake."""
+    b = tmp_path / "b"
+    b.mkdir()
+    (b / MANIFEST_FILE).write_bytes(payload)
+
+    got = read_manifest(b)
+    assert isinstance(got, Manifest)
+    assert got.error, "a damaged manifest read clean"
+
+
+def test_damaged_is_not_the_same_as_absent(tmp_path):
+    """INV-5. Silently folding "cannot be read" into "never announced" would
+    hide the one case somebody has to go and fix."""
+    absent = tmp_path / "absent"
+    absent.mkdir()
+    damaged = tmp_path / "damaged"
+    damaged.mkdir()
+    (damaged / MANIFEST_FILE).write_text("{oops")
+
+    assert read_manifest(absent) is None
+    assert read_manifest(damaged).error
+
+
+def test_a_manifest_the_module_did_not_write_still_reads(tmp_path):
+    """Hand-written is a supported input: the file is plain JSON in a folder the
+    operator owns, and half the point is that a booth is just a directory. Only
+    `handle` is required; everything else has a default."""
+    b = tmp_path / "b"
+    b.mkdir()
+    (b / MANIFEST_FILE).write_text(json.dumps({"handle": "shutter-dev"}))
+
+    got = read_manifest(b)
+    assert got.handle == "shutter-dev" and got.error is None
+    assert got.title == "b"
+    assert got.why == ""
+
+
+# ---- slice 3: re-announcement (INV-3) ---------------------------------------
+
+
+def test_re_announcing_preserves_created(tmp_path):
+    """INV-3. `created` is when the booth APPEARED. Saying something more about
+    it later is not a second appearance, and a `booth add` on an existing booth
+    is the common case — the poster adds the second batch and sharpens the why."""
+    b = tmp_path / "b"
+    b.mkdir()
+    first = write_manifest(b, "booth-dev", why="first pass")
+    second = write_manifest(b, "booth-dev", why="second pass, sharper")
+
+    assert second.created == first.created
+    assert second.why == "second pass, sharper"
+
+
+def test_re_announcing_over_a_damaged_file_does_not_inherit_its_created(tmp_path):
+    """A `created` that cannot be read back is replaced rather than guessed at.
+    The alternative is a stamp that is silently wrong, which is worse than one
+    that is silently new."""
+    b = tmp_path / "b"
+    b.mkdir()
+    (b / MANIFEST_FILE).write_text("{not json")
+
+    m = write_manifest(b, "booth-dev", why="rescued")
+    assert m.created and m.error is None
+    assert read_manifest(b).why == "rescued"
+
+
+# ---- slice 4: the write is atomic, and invisible to every listing -----------
+
+
+def test_the_write_leaves_no_temp_file(tmp_path):
+    """Half of the atomic-write promise, and the weaker half — see
+    `test_the_write_replaces_rather_than_truncating` for the part that actually
+    discriminates. Kept because a leaked `.tmp` is its own small defect: it
+    would sit in the booth forever and, unlike the manifest, nothing would ever
+    overwrite it."""
+    b = tmp_path / "b"
+    b.mkdir()
+    write_manifest(b, "booth-dev", why="x")
+    assert not list(b.glob("*.tmp")), "a temp file survived the write"
+
+
+def test_a_manifest_is_not_an_item(tmp_path):
+    """The whole integration story: it is a DOTFILE, so the existing
+    `startswith('.')` skip in `booth_items` already keeps it out of tiles,
+    counts and zips. No new exclusion rule anywhere. Asserted rather than
+    assumed, because the claim is load-bearing for the contract's scope."""
+    from booth.app import zip_booth
+    from booth.items import booth_items
+
+    b = tmp_path / "b"
+    b.mkdir()
+    (b / "a.txt").write_text("real content")
+    write_manifest(b, "booth-dev", why="x")
+
+    assert [i.rel for i in booth_items(b)] == ["a.txt"]
+    assert MANIFEST_FILE not in zip_booth(b).decode("latin-1")
+
+
+def test_announcing_is_activity(tmp_path):
+    """A manifest is a dotfile but not a `.lock` dotfile, so `_newest_mtime`
+    counts it. Creating or re-announcing a booth resets its TTL, which is right:
+    both are somebody touching it. The lock exemption added in v0.2.2 is for
+    machinery a READ path creates; this is a deliberate write."""
+    from booth.app import booth_age_seconds
+
+    b = tmp_path / "b"
+    b.mkdir()
+    old = 1_000_000_000
+    os.utime(b, (old, old))
+
+    write_manifest(b, "booth-dev", why="look at this")
+    assert booth_age_seconds(b, now=old + 90_000) < 86_400
+
+
+def test_stdlib_only():
+    """INV-4, and the reason this module exists separately from anything that
+    imports a third-party package. `scripts/booth` imports it under the system
+    python3 with NO venv, through a `python3 -c` heredoc no AST extractor can
+    see. It must also not import `booth.*`: a cross-import between two
+    stdlib-only modules is a second way for the invariant to break."""
+    src = pathlib.Path(__file__).parent.parent / "booth" / "manifest.py"
+    roots = set()
+    for node in ast.walk(ast.parse(src.read_text())):
+        if isinstance(node, ast.Import):
+            roots.update(a.name.split(".")[0] for a in node.names)
+        elif isinstance(node, ast.ImportFrom):
+            # A RELATIVE import (`from . import marks`) carries no module root
+            # and used to pass this walk unseen — which matters more here than
+            # in the shared copy, because this module forbids sibling imports
+            # outright. Recorded as `booth` so the assertion below catches it.
+            roots.add("booth" if node.level else
+                      (node.module or "").split(".")[0])
+    assert not (roots - set(sys.stdlib_module_names)), (
+        f"booth/manifest.py imports outside the stdlib: "
+        f"{sorted(roots - set(sys.stdlib_module_names))}"
+    )
+
+
+# ---- slice 5: what the operator actually sees -------------------------------
+
+
+@pytest.fixture
+def client(tmp_path):
+    from fastapi.testclient import TestClient
+
+    from booth.app import create_app
+    return TestClient(create_app(tmp_path, ttl_hours=24, start_sweeper=False)), tmp_path
+
+
+def _booth(data, name, *, kept=False):
+    b = data / name
+    b.mkdir()
+    (b / "a.txt").write_text("content")
+    if kept:
+        (b / ".forever").touch()
+    return b
+
+
+@pytest.mark.parametrize("kept", [False, True], ids=["ephemeral", "kept"])
+def test_the_index_card_carries_the_announcement(client, kept):
+    """BOTH LANES. Kept boards render first and are a separate block in
+    index.html, so patching only the ephemeral lane would leave the 15 kept
+    booths — the durable, most-looked-at ones — with exactly the defect this
+    unit closes. Same lesson as the `blurtoggle` macro: three branches, one
+    definition; here it is two lanes and one rule."""
+    c, data = client
+    b = _booth(data, "r18-ab", kept=kept)
+    write_manifest(b, "booth-dev", why="pick the winning denoiser")
+
+    html = c.get("/").text
+    assert "booth-dev" in html
+    assert "pick the winning denoiser" in html
+
+
+@pytest.mark.parametrize("kept", [False, True], ids=["ephemeral", "kept"])
+def test_a_booth_that_never_spoke_up_is_marked(client, kept):
+    """All 26 live booths are in this state, and rsync keeps making more. The
+    marker is what makes the convention adoptable at all: the link board rotted
+    to 69% precisely because nothing ever showed which rows were dead.
+
+    ASSERTED ON THE CLASS, not on the word, and the test is named around it.
+    `pytest`'s `tmp_path` is derived from the TEST NAME and the index renders
+    `data_dir` in its empty-state hint — so a test called
+    `test_an_unannounced_booth_says_so` put the literal string "unannounced"
+    into the page and passed against a template that did not yet exist. A
+    structural hook cannot be spelled by accident — though it has to be the
+    rendered ELEMENT and not the bare class, since base.html ships a
+    `.prov-none{...}` rule into the very same page."""
+    c, data = client
+    _booth(data, "quiet", kept=kept)
+
+    html = c.get("/").text
+    assert 'class="prov prov-none"' in html
+    assert "unannounced" in html
+
+
+def test_a_damaged_manifest_reads_differently_from_an_absent_one(client):
+    """INV-5 on the surface the operator looks at, not just in the reader."""
+    c, data = client
+    b = _booth(data, "damaged")
+    (b / MANIFEST_FILE).write_text("{oops")
+
+    html = c.get("/").text
+    assert 'class="prov prov-broken"' in html
+    assert "unreadable" in html
+    assert c.get("/b/damaged/").status_code == 200
+
+
+def test_an_announced_booth_with_no_why_shows_only_its_handle(client):
+    """`booth new x` with no --why is legal and common. The card shows who made
+    it and does not invent a purpose or leave a dangling separator."""
+    c, data = client
+    b = _booth(data, "scratch")
+    write_manifest(b, "booth-dev")
+
+    html = c.get("/").text
+    assert "booth-dev" in html
+    assert 'class="prov prov-none"' not in html
+
+
+def test_the_booth_page_header_carries_it_too(client):
+    """Deliberate scope, not creep: a booth URL handed to the operator lands
+    HERE, never on the index. Job 5 is 'operator, look at this', so the page he
+    actually opens is where the answer has to be."""
+    c, data = client
+    b = _booth(data, "r18-ab")
+    write_manifest(b, "booth-dev", why="pick the winning denoiser")
+
+    html = c.get("/b/r18-ab/").text
+    assert "booth-dev" in html
+    assert "pick the winning denoiser" in html
+
+
+def test_a_poisoned_manifest_cannot_take_down_the_index(client):
+    """The v0.2.2 lesson, asserted for the new reader before it can repeat:
+    `list_booths` touches every booth on every page load, so one bad file must
+    cost that booth's provenance and nothing else."""
+    c, data = client
+    _booth(data, "good")
+    bad = _booth(data, "bad")
+    (bad / MANIFEST_FILE).write_bytes(b"\xff\xfe not utf-8 at all")
+
+    assert c.get("/").status_code == 200
+    assert c.get("/healthz").status_code == 200
+
+
+def test_a_pickup_booth_announces_itself_as_the_booths_own(client):
+    """No exemption list. A booth the service made says the service made it,
+    which is true — and it keeps the rule to one line: a booth with no manifest
+    is unannounced."""
+    c, data = client
+    r = c.post("/upload", files=[("files", ("a.txt", b"hello", "text/plain"))],
+               follow_redirects=False)
+    assert r.status_code in (200, 303)
+
+    booth = next(p for p in data.iterdir() if p.is_dir())
+    got = read_manifest(booth)
+    assert got is not None and got.handle == "booth"
+    assert 'class="prov prov-none"' not in c.get("/").text
+
+
+# ---- findings from the cross-frontier CODE-REVIEW panel, 2026-09-22 ----------
+#
+# Heid panel (thread 01M341E9XAPZEFBSPK9HPGAM0S). Four arms, artifact-only.
+# The round found ZERO drift in the strict sense and landed its weight one layer
+# down, in test strength: five of the ten adopted findings are tests of mine
+# that pass on the regression they exist to catch.
+
+
+def test_the_read_survives_a_document_no_one_can_parse(tmp_path):
+    """INV-2 said "never raises" and named a 4 GB file as a tested case. It was
+    not tested, and it did not hold: `except ValueError` catches a truncated
+    document, but `json.loads` on deeply nested input raises RecursionError,
+    which is not a ValueError and is not an OSError either.
+
+    `list_booths` calls this once per booth on every index load, so the one
+    file costs the whole front page — the exact outage shape the invariant
+    cites as its reason for existing. Three of four arms reached it
+    independently; the eight-payload parametrize above has no size or depth
+    case, so the hole stayed green.
+    """
+    b = tmp_path / "b"
+    b.mkdir()
+    (b / MANIFEST_FILE).write_text("[" * 200_000 + "]" * 200_000)
+
+    got = read_manifest(b)
+    assert isinstance(got, Manifest) and got.error
+
+
+def test_the_read_refuses_a_document_too_large_to_be_a_manifest(tmp_path):
+    """The other half of INV-2's named case. A manifest is four short fields;
+    anything approaching a megabyte is not one, and reading it into memory to
+    discover that is the wrong order of operations. Bounded BEFORE the read, so
+    the size is checked by `stat` rather than survived."""
+    from booth.manifest import MANIFEST_MAX_BYTES
+
+    b = tmp_path / "b"
+    b.mkdir()
+    (b / MANIFEST_FILE).write_text('{"handle": "x", "why": "' +
+                                   "y" * (MANIFEST_MAX_BYTES + 100) + '"}')
+
+    got = read_manifest(b)
+    assert isinstance(got, Manifest) and got.error
+    assert "too large" in got.error
+
+
+def test_a_hostile_directory_name_does_not_reach_the_record_raw(tmp_path):
+    """`_one_line(title, TITLE_MAX) or booth.name` — the FALLBACK skips the
+    normalization the explicit value gets. A directory name may legally carry a
+    newline on POSIX and may be 255 bytes, and either lands in a card's
+    sub-line. Same shape on the read path's fallback."""
+    # 200-odd bytes, under the filesystem's own 255 limit but well over
+    # TITLE_MAX — and a newline, which POSIX permits in a filename.
+    name = "we" + "i" * 200 + "rd\nname"
+    b = tmp_path / name
+    b.mkdir()
+
+    m = write_manifest(b, "booth-dev")
+    assert "\n" not in m.title and len(m.title) <= 120
+    assert "\n" not in read_manifest(b).title
+
+
+def test_the_write_replaces_rather_than_truncating(tmp_path):
+    """The previous version of this test asserted only that no `*.tmp` file
+    survived — which a plain `write_text` passes, since it leaves no temp file
+    either. All four arms said so, and they were right.
+
+    THE INODE IS THE DISCRIMINATOR. `os.replace` publishes a different file over
+    the old name, so the inode changes; truncate-and-rewrite keeps it. That is
+    also exactly why the promise holds for a concurrent reader: it either has
+    the old inode, intact, or opens the new one, complete. A test of the
+    mechanism rather than of its litter.
+
+    (An earlier draft spied on `os.open` to prove the published path was never
+    opened for writing. It passed — vacuously. `Path.write_text` reaches the
+    syscall through `io.open` in C and never touches the Python-level
+    `os.open`, so the spy could not have fired either way. Recorded because
+    writing a second vacuous test while fixing the first is the failure mode
+    this whole round is about.)
+    """
+    b = tmp_path / "b"
+    b.mkdir()
+    published = b / MANIFEST_FILE
+
+    write_manifest(b, "booth-dev", why="first")
+    first_inode = published.stat().st_ino
+    write_manifest(b, "booth-dev", why="second")
+
+    assert published.stat().st_ino != first_inode, (
+        "the manifest was rewritten in place, not replaced"
+    )
+    assert read_manifest(b).why == "second"
+
+
+def test_the_temp_file_is_not_a_name_two_writers_share(tmp_path):
+    """Every writer derived the same `.booth.json.tmp`. Two `booth add` calls on
+    one booth could then interleave through a stale descriptor into the
+    published path — the atomic-write promise is that READERS never see a
+    partial file, and it says nothing about two writers sharing a scratch name.
+    Marks are protected from this by their flock; the manifest has none."""
+    b = tmp_path / "b"
+    b.mkdir()
+    seen = set()
+    for i in range(5):
+        write_manifest(b, "booth-dev", why=f"pass {i}")
+        seen.update(p.name for p in b.iterdir() if p.name != MANIFEST_FILE)
+    assert not seen, f"left temp files behind: {sorted(seen)}"
+
+    from booth.manifest import _temp_path
+    names = {_temp_path(b).name for _ in range(20)}
+    assert len(names) > 1, "every writer derives the same temp name"
+
+
+def test_a_bare_re_announce_does_not_wipe_the_why(tmp_path):
+    """THE WORKFLOW IS `new --why` THEN `add`. Omitted flags meant empty
+    strings, and empty strings overwrote — so the second command silently
+    erased the sentence the first one existed to record, on the single most
+    common sequence this feature has.
+
+    Two arms of the paraphrase panel predicted it from the contract's wording
+    alone ("gains a manifest with no why" does not distinguish a first write
+    from a re-announce with the flags omitted). Every test I wrote passed
+    `--why` on both calls, so none of them could see it.
+
+    Omitted now means UNCHANGED; only a value that was actually supplied
+    overwrites, and an explicit empty string still clears.
+    """
+    b = tmp_path / "b"
+    b.mkdir()
+    write_manifest(b, "booth-dev", title="R18 A/B", why="pick the denoiser")
+
+    write_manifest(b, "booth-dev")                       # a bare `booth add`
+    kept = read_manifest(b)
+    assert kept.why == "pick the denoiser", "a bare re-announce wiped the why"
+    assert kept.title == "R18 A/B"
+
+    write_manifest(b, "booth-dev", why="sharper")        # supplied: overwrites
+    assert read_manifest(b).why == "sharper"
+
+    write_manifest(b, "booth-dev", why="")               # explicit: clears
+    assert read_manifest(b).why == ""
+
+
+def test_re_announcing_preserves_a_created_from_before_this_second(tmp_path):
+    """`_now()` is whole-second resolution, so two `write_manifest` calls in a
+    row share a timestamp and the old preservation test passed even against an
+    implementation that regenerated `created` every time. Three of four arms
+    caught it. Seed a stamp that could not have come from now()."""
+    b = tmp_path / "b"
+    b.mkdir()
+    (b / MANIFEST_FILE).write_text(json.dumps({
+        "handle": "booth-dev", "title": "b", "why": "first",
+        "created": "2019-03-04T11:22:33-08:00",
+    }))
+
+    assert write_manifest(b, "booth-dev", why="second").created == \
+        "2019-03-04T11:22:33-08:00"
+
+
+def test_only_the_manifest_module_opens_the_manifest(tmp_path):
+    """INV-1, which had no guard anywhere. One resolver is only one resolver
+    while nothing else learns the filename."""
+    root = pathlib.Path(__file__).parent.parent
+    offenders = []
+    for src in sorted((root / "booth").glob("*.py")):
+        if src.name == "manifest.py":
+            continue
+        tree = ast.parse(src.read_text())
+        # STRING CONSTANTS, not raw text. A comment naming the file is prose
+        # about the design and harms nothing — the first version of this test
+        # scanned the whole source and went red on a comment explaining why a
+        # leaked `.booth.json.<hex>.tmp` keeps a booth alive. The invariant is
+        # about code that knows the filename, so ask the code.
+        docstrings = set()
+        for node in ast.walk(tree):
+            if isinstance(node, (ast.Module, ast.ClassDef,
+                                 ast.FunctionDef, ast.AsyncFunctionDef)):
+                body = getattr(node, "body", None)
+                if body and isinstance(body[0], ast.Expr) and \
+                        isinstance(body[0].value, ast.Constant):
+                    docstrings.add(id(body[0].value))
+        for node in ast.walk(tree):
+            if (isinstance(node, ast.Constant) and isinstance(node.value, str)
+                    and id(node) not in docstrings and ".booth.json" in node.value):
+                offenders.append(f"{src.name}:{node.lineno}")
+    assert not offenders, f"{offenders} name the manifest file in code"
+
+
+def test_announcing_is_activity_via_the_manifest_file_itself(tmp_path):
+    """The previous version could not fail. Writing the manifest creates a
+    directory entry, which bumps the DIRECTORY's mtime, so the booth read as
+    fresh whether or not `_newest_mtime` counted the manifest at all — a test
+    of the side effect rather than of the thing.
+
+    Put the directory's clock back afterwards, leaving the manifest's own mtime
+    as the only thing that can keep the booth alive."""
+    import os
+
+    from booth.app import booth_age_seconds
+
+    b = tmp_path / "b"
+    b.mkdir()
+    old = 1_000_000_000
+    os.utime(b, (old, old))
+    write_manifest(b, "booth-dev", why="look at this")
+    os.utime(b, (old, old))                  # only the file can save it now
+
+    assert booth_age_seconds(b, now=old + 90_000) < 86_400
+
+
+def test_the_booth_header_marks_an_unannounced_booth_too(client):
+    """The negative states were asserted on `/` only, so a header that rendered
+    provenance for clean manifests and nothing for the other two would have
+    passed the whole suite."""
+    c, data = client
+    _booth(data, "quiet")
+    damaged = _booth(data, "damaged")
+    (damaged / MANIFEST_FILE).write_text("{oops")
+
+    assert 'class="prov prov-none"' in c.get("/b/quiet/").text
+    assert 'class="prov prov-broken"' in c.get("/b/damaged/").text
+
+
+def test_the_title_reaches_a_surface(client):
+    """`--title` promised a display name and nothing rendered it — 4/4 on the
+    paraphrase panel, independently the top-ranked flag of that round. It lands
+    on the booth page heading, where there is room for it; the INDEX card keeps
+    the directory name, because that is the identity the operator navigates and
+    refers to positionally."""
+    c, data = client
+    b = _booth(data, "r18-ab")
+    write_manifest(b, "booth-dev", title="R18 A/B — denoiser bakeoff", why="w")
+
+    page = c.get("/b/r18-ab/").text
+    assert "R18 A/B — denoiser bakeoff" in page
+    assert "r18-ab" in page, "the directory name stopped being visible"
+
+
+# ---- findings from the cross-frontier BUG-HUNT panel, 2026-09-22 -------------
+#
+# Heid panel (thread 01M343SXX27Z47C3STXXRC7M42). Four arms, artifact-only,
+# diff-scoped. The strongest finding is one the SIZE CAP ITSELF opened.
+
+
+def test_a_reader_never_blocks_on_a_file_that_is_not_a_file(tmp_path):
+    """`stat` reports size 0 for a FIFO, so it sails under the byte cap — and
+    then `read_text` blocks in `read` with no EOF, so the `except` never runs
+    and the call never returns. `list_booths` reads every booth on every `GET /`
+    and `/healthz`, so ONE such file stalls the front page for the whole service,
+    with no error and no recovery short of a restart.
+
+    A symlink to `/dev/zero` is the same hole with unbounded allocation instead
+    of a hang: `st_size` is 0 there too.
+
+    Two of four arms reached it independently. The bound added an hour earlier
+    is what made it reachable — `st_size` answers a different question than
+    "can this be read", and a cap that trusts it inherits the difference.
+    """
+    import os
+    import signal
+
+    b = tmp_path / "b"
+    b.mkdir()
+    os.mkfifo(b / MANIFEST_FILE)
+
+    # ⚠ ALARMED. Without this the RED state of this test does not fail, it HANGS
+    # — which is the defect itself, and is also useless as a signal: a suite that
+    # stops is indistinguishable from a suite that is slow. Five seconds is a
+    # thousand times the budget a read of a four-field file should need.
+    def _timeout(signum, frame):
+        raise AssertionError("read_manifest blocked on a FIFO and never returned")
+
+    old_handler = signal.signal(signal.SIGALRM, _timeout)
+    signal.alarm(5)
+    try:
+        got = read_manifest(b)
+    finally:
+        signal.alarm(0)
+        signal.signal(signal.SIGALRM, old_handler)
+
+    assert isinstance(got, Manifest) and got.error
+    assert "regular file" in got.error
+
+
+def test_a_damaged_manifest_is_kept_when_it_is_replaced(tmp_path):
+    """4/4, and it contradicted this repo's own doctrine. Marks made the rule
+    explicit in v0.2.1 — reads stay lenient, writes go strict, damaged bytes
+    STAY ON DISK — and the manifest's write replaced them outright.
+
+    The sharpest leg: a file that fails on ONE field still holds the others.
+    `{"handle": 7, "why": "the thing I wanted you to look at"}` reads as broken
+    and used to be destroyed whole, taking a `why` the re-announcer may not have
+    kept anywhere.
+
+    Quarantined rather than refused: refusing would fail `booth add` and lose
+    the files it was copying, which is the worse trade. One fixed-name
+    quarantine, so this cannot accumulate.
+    """
+    from booth.manifest import QUARANTINE_FILE
+
+    b = tmp_path / "b"
+    b.mkdir()
+    damaged = json.dumps({"handle": 7, "why": "the thing I wanted you to see"})
+    (b / MANIFEST_FILE).write_text(damaged)
+
+    write_manifest(b, "booth-dev", why="rescued")
+
+    assert read_manifest(b).why == "rescued"
+    assert (b / QUARANTINE_FILE).read_text() == damaged, "the damaged bytes were destroyed"
+
+
+def test_a_broken_record_normalizes_the_directory_name_too(tmp_path):
+    """The third fallback. `write_manifest`'s and `read_manifest`'s were fixed
+    in the previous round and `_broken`'s was missed — same raw `booth.name`,
+    same card sub-line, same newline."""
+    b = tmp_path / ("wei" + "i" * 200 + "rd\nname")
+    b.mkdir()
+    (b / MANIFEST_FILE).write_text("{oops")
+
+    got = read_manifest(b)
+    assert got.error and "\n" not in got.title and len(got.title) <= 120
+
+
+def test_an_identical_re_announce_does_not_touch_the_booth(tmp_path):
+    """Marks learned this in v0.2.0: a write that changes nothing is not
+    activity and must not reset a booth's TTL. The manifest wrote
+    unconditionally, so `booth add` on an unchanged booth kept a dead one alive
+    — and `booth link` does it on every single post to the standing board."""
+    import os
+
+    b = tmp_path / "b"
+    b.mkdir()
+    write_manifest(b, "booth-dev", why="x")
+    path = b / MANIFEST_FILE
+    os.utime(path, (1_000_000_000, 1_000_000_000))
+    os.utime(b, (1_000_000_000, 1_000_000_000))
+    before = path.stat().st_mtime
+
+    write_manifest(b, "booth-dev", why="x")          # identical
+
+    assert path.stat().st_mtime == before, "an identical re-announce rewrote the file"
+
+
+def test_a_failed_write_leaves_no_temp_file_behind(tmp_path):
+    """The unique temp name fixed a cross-writer hazard and created a litter
+    one: a fixed name is overwritten by the next writer, a random one is not.
+    And `.booth.json.<hex>.tmp` is NOT a `.lock`, so `_newest_mtime` counts it —
+    an orphaned temp would keep a dead booth alive forever."""
+    import os
+
+    b = tmp_path / "b"
+    b.mkdir()
+    real_replace = os.replace
+
+    def boom(src, dst, *a, **kw):
+        raise OSError("no space left on device")
+
+    os.replace = boom
+    try:
+        with pytest.raises(OSError):
+            write_manifest(b, "booth-dev", why="x")
+    finally:
+        os.replace = real_replace
+
+    assert not list(b.glob("*.tmp")), f"orphaned temp: {list(b.glob('*.tmp'))}"

tests/test_marks.py

@@ -7,6 +7,8 @@ See docs/contracts/u2_marks.contract.md.
 """
 import ast
 import json
+import re
+import tomllib
 import pathlib
 import sys
 
@@ -276,19 +278,31 @@ def test_as_dict_round_trips_through_json(tmp_path):
 # ---- the stdlib-only invariant (INV-5) --------------------------------------
 
 
-@pytest.mark.parametrize("module", ["marks", "asks", "links"])
+@pytest.mark.parametrize("module", ["marks", "asks", "links", "manifest", "benches", "__init__"])
 def test_stdlib_only(module):
     """INV-5. scripts/booth imports these under the system python3 with NO venv,
     through a `python3 -c` heredoc that no AST extractor can see — so nothing
     but this test stands between a casual third-party import and `booth ask`
     breaking on every fleet host."""
+    # `manifest` also carries a stricter copy in tests/test_manifest.py, which
+    # additionally forbids importing `booth.*` — a cross-import between two
+    # stdlib-only modules is a second way for this invariant to break.
     src = pathlib.Path(__file__).parent.parent / "booth" / f"{module}.py"
     tree = ast.parse(src.read_text())
     roots = set()
     for node in ast.walk(tree):
         if isinstance(node, ast.Import):
             roots.update(a.name.split(".")[0] for a in node.names)
-        elif isinstance(node, ast.ImportFrom) and node.level == 0 and node.module:
+        elif isinstance(node, ast.ImportFrom):
+            # `node.level > 0` is a RELATIVE import (`from . import marks`),
+            # which has no `module` root to inspect and used to slip through
+            # this walk entirely. It cannot reach outside the package, so it is
+            # stdlib-safe by construction — but it is recorded rather than
+            # ignored, because `manifest.py` additionally forbids importing a
+            # sibling and its own test needs to see one.
+            if node.level:
+                roots.add("booth")
+            elif node.module:
                 roots.add(node.module.split(".")[0])
     outside = {r for r in roots if r != "booth" and r not in sys.stdlib_module_names}
     assert not outside, f"booth/{module}.py imports non-stdlib: {sorted(outside)}"
@@ -881,3 +895,566 @@ def test_a_corrupt_marks_file_gives_the_browser_a_409_not_a_500(client):
     # the page still renders, so the operator can see the booth at all
     assert c.get("/b/b/").status_code == 200
     assert c.get("/b/b/marks.json").status_code == 200
+
+
+# ---- findings from the cross-frontier BUG-HUNT panel, 2026-09-22 -------------
+#
+# Heid panel (thread 01M33XEC1H0298C0D968FWBN7A). Four arms, artifact-only,
+# diff-scoped. The headline was 4/4 convergent and none of it had a guard: the
+# panel's own mutation tables showed the lock lifecycle SURVIVED every existing
+# test, because `test_a_no_op_write_does_not_touch_the_booth` asserts only that
+# `.marks.json` is absent and never looks at the lock or at the clock the
+# sweeper actually reads.
+
+
+def test_the_lock_file_is_never_unlinked(tmp_path):
+    """The lock must outlive the operation that created it.
+
+    `flock` binds to an INODE, not to a path. Unlinking `.marks.lock` while a
+    second writer is blocked on it leaves that writer holding an exclusive lock
+    on a deleted inode — and the next writer along creates a FRESH lock file and
+    takes it immediately. Two processes then run the read-modify-write
+    concurrently and the later `os.replace` drops the earlier one's mark, with
+    no error anywhere. Both of them obeyed the protocol.
+
+    The cleanup existed to keep a no-op from leaving a lock file as its only
+    trace. That is a tidiness goal, and it bought a lost-update race.
+    """
+    from booth.marks import MARKS_LOCK, set_flag
+
+    booth = tmp_path / "b"
+    booth.mkdir()
+    assert set_flag(booth, "ghost.png", False) is None      # a no-op
+    assert (booth / MARKS_LOCK).exists(), "the no-op path unlinked the lock file"
+
+
+def test_a_no_op_does_not_reset_the_ttl_clock(tmp_path):
+    """The property the no-op guard actually exists for, asserted against the
+    clock the sweeper reads instead of against one file's absence.
+
+    Creating or removing a directory entry bumps the DIRECTORY's mtime, and
+    `_newest_mtime` seeds from exactly that. So `touch` + `unlink` of the lock
+    reset the booth's age to zero while leaving no trace behind — the comment on
+    the create-only guard reasons about the lock FILE's mtime and misses that
+    the directory moved underneath it. Repeated, it kept a dead booth alive
+    forever, which is the precise outcome the guard was written to prevent.
+    """
+    import os
+
+    from booth.app import booth_age_seconds
+    from booth.marks import delete_mark, set_flag
+
+    booth = tmp_path / "b"
+    booth.mkdir()
+    old = 1_000_000_000
+    os.utime(booth, (old, old))
+
+    set_flag(booth, "ghost.png", False)     # no-op: never flagged
+    delete_mark(booth, "nothing")           # no-op: no such mark
+
+    age = booth_age_seconds(booth, now=old + 90_000)
+    assert age > 86_400, f"a no-op reset the TTL clock (age fell to {age:.0f}s)"
+
+
+def test_a_real_mark_still_resets_the_ttl_clock(tmp_path):
+    """The other half of the same rule, so the fix cannot overshoot into
+    'marking is never activity'. Marking IS activity and must reset the clock;
+    only a write that changes nothing must not."""
+    import os
+
+    from booth.app import booth_age_seconds
+    from booth.marks import set_flag
+
+    booth = tmp_path / "b"
+    booth.mkdir()
+    old = 1_000_000_000
+    os.utime(booth, (old, old))
+
+    set_flag(booth, "a.png", True)          # a real mark
+
+    assert booth_age_seconds(booth, now=old + 90_000) < 86_400
+
+
+def test_a_non_string_note_text_does_not_crash_the_read(tmp_path):
+    """`_clean_text` did `(text or "").replace(...)`, so a stored `text` that is
+    valid JSON but not a string raised AttributeError out of the READ path.
+
+    That is not a marks bug, it is an INDEX bug: `list_booths` reads every
+    booth's marks on every page load, so one poisoned file took down `/` and
+    `/healthz` for all 25 booths. The module's stated posture is that a mark it
+    cannot read renders as broken, never as a 500.
+    """
+    booth = tmp_path / "b"
+    booth.mkdir()
+    (booth / MARKS_FILE).write_text(json.dumps({
+        "version": 1,
+        "marks": [{"id": "n1", "shape": "note", "text": 7,
+                   "created": "2026-09-21T00:00:00+00:00"}],
+    }))
+
+    marks = marks_for(booth)
+    assert len(marks) == 1
+    assert marks[0].error, "a poisoned note read clean instead of reading broken"
+
+
+def test_a_non_string_created_does_not_crash_the_sort(tmp_path):
+    """`marks_for` sorts on `(created, id)`. A stored `created` of the wrong type
+    made that comparison raise TypeError — same blast radius as the note above,
+    reached through the sort rather than through hydration."""
+    booth = tmp_path / "b"
+    booth.mkdir()
+    (booth / MARKS_FILE).write_text(json.dumps({
+        "version": 1,
+        "marks": [
+            {"id": "a", "shape": "note", "text": "fine",
+             "created": "2026-09-21T00:00:00+00:00"},
+            {"id": "b", "shape": "note", "text": "also fine", "created": 17},
+        ],
+    }))
+
+    marks = marks_for(booth)
+    assert len(marks) == 2
+    # An unreadable mark loses its `created` and so sorts FIRST — the stated
+    # rule is `("", id)` against `(created, id)`. A mark nobody can read is the
+    # one that wants looking at, and the alternative is it landing at an
+    # arbitrary position in the middle of the panel.
+    assert [m.id for m in marks] == ["b", "a"]
+    assert marks[0].error and not marks[1].error
+
+
+def test_legacy_import_order_survives_same_second_mtimes(tmp_path):
+    """ROADMAP states the legacy import's order is `(mtime, name)`. It was
+    stamping `created` at whole-second resolution, so two sidecars written in
+    the same second lost the fractional part that distinguished them and
+    `marks_for`'s `(created, id)` tie-break silently re-sorted them into
+    alphabetical order — reversing the pair the importer had just ordered.
+
+    Deterministic order is a v1 invariant precisely because the operator refers
+    to things positionally. An order that is stated and not kept is worse than
+    one that was never claimed.
+    """
+    import os
+
+    from booth.marks import import_legacy_asks
+
+    booth = tmp_path / "b"
+    booth.mkdir()
+    for stem in ("zeta", "alpha"):
+        (booth / f"{stem}{ASK_SUFFIX}").write_text(json.dumps(_single()))
+    # Same whole second, different fractions: `zeta` is OLDER and must come first.
+    os.utime(booth / f"zeta{ASK_SUFFIX}", (1_700_000_000.10, 1_700_000_000.10))
+    os.utime(booth / f"alpha{ASK_SUFFIX}", (1_700_000_000.90, 1_700_000_000.90))
+
+    imported = [m.id for m in import_legacy_asks(booth)]
+    assert imported == ["zeta", "alpha"], "the importer's own order is wrong"
+    assert [m.id for m in marks_for(booth)] == imported, (
+        "the read path re-sorted what the importer ordered"
+    )
+
+
+def test_the_index_survives_a_poisoned_marks_file(client):
+    """The blast radius, asserted where it actually hurts.
+
+    `list_booths` reads every booth's marks on every index load and `/healthz`
+    does the same. One hand-edited or foreign-written `.marks.json` therefore
+    took down the front page for all 25 booths — the single-booth failure the
+    lenient reader exists to contain, escaping the booth it belongs to.
+    """
+    c, data = client
+    good = data / "good"
+    good.mkdir()
+    _png(good / "a.png")
+    bad = data / "bad"
+    bad.mkdir()
+    (bad / MARKS_FILE).write_text(json.dumps({
+        "version": 1,
+        "marks": [{"id": "n1", "shape": "note", "text": {"oops": True}, "created": 3}],
+    }))
+
+    assert c.get("/").status_code == 200
+    assert c.get("/healthz").status_code == 200
+    assert c.get("/b/bad/").status_code == 200
+
+
+def test_answer_treats_a_non_string_notes_field_as_no_notes(client):
+    """`booth_note` guards `text` with `isinstance(..., str)`; `booth_answer`
+    passed `notes` straight to `_clean_notes`, which calls `.replace` on it. A
+    multipart FILE part named `notes` is a str to nobody, so the route 500'd on
+    hostile-but-legal input where its sibling handled the same class of value.
+
+    Both routes now read the field the same way: a value that is not text is no
+    value. The CHOICE is the judgment and it still lands — throwing the whole
+    answer away over a junk optional field would be the wrong trade."""
+    c, data = client
+    b = data / "b"
+    b.mkdir()
+    declare_pick(b, "winner", _single())
+
+    r = c.post(
+        "/b/b/answer",
+        data={"ask": "winner", "choice": "A — baseline"},
+        files={"notes": ("n.txt", b"surprise", "text/plain")},
+        follow_redirects=False,
+    )
+    assert r.status_code == 303
+    mark = next(m for m in marks_for(b) if m.id == "winner")
+    assert mark.answer["choice"] == "A — baseline"
+    assert not mark.answer.get("notes")
+
+
+def test_an_inline_doc_tile_offers_a_note_control(client):
+    """Three item branches, two of them call `marknotes`. The doc branch got the
+    flag button and not the note field, so the operator could point at a report
+    and not write down why — on the one item kind whose whole purpose is prose.
+
+    This is the exact failure the `blurtoggle` macro comment names ("patched two
+    of three"), recurring on the macro that was written to prevent it.
+    """
+    c, data = client
+    b = data / "b"
+    b.mkdir()
+    (b / "report.md").write_text("# report\n\nprose here\n")
+
+    html = c.get("/b/b/").text
+    assert 'value="report.md"' in html, "the doc tile has no mark controls at all"
+    # `marknotes`' add-field, which only that macro emits. The booth-level panel
+    # has its own note form, so the presence of /note on the page proves nothing.
+    assert 'placeholder="a note on this item"' in html, (
+        "an inline doc tile has no way to add a note"
+    )
+
+
+def test_the_marks_panel_survives_a_booth_that_also_has_a_link_board(client):
+    """The board booth renders as a board instead of a gallery, which is right —
+    but the suppression was unconditional, so a pick declared on a booth that
+    happens to carry a `links.md` had no form to answer it and no way to say so."""
+    c, data = client
+    b = data / "b"
+    b.mkdir()
+    (b / "links.md").write_text("- [a thing](http://example.invalid) <sub>· who · when</sub>\n")
+    declare_pick(b, "winner", _single())
+
+    html = c.get("/b/b/").text
+    assert "Which render wins?" in html, "a pick on a board booth was unanswerable"
+
+
+def test_the_zoom_view_does_not_navigate_away_from_a_note_being_typed(client):
+    """The viewer's arrow keys move between images and Escape goes back. The
+    note textarea landed in the same page, and the handler is on `document`, so
+    an arrow key meant for the caret threw away the draft instead of moving it.
+
+    Asserted structurally: the handler must bail on events from an editable
+    target. There is no browser in this suite, and a guard nobody can test is
+    exactly how this shipped."""
+    c, data = client
+    b = data / "b"
+    b.mkdir()
+    _png(b / "a.png")
+
+    js = c.get("/b/b/view?f=a.png").text
+    assert "isEditable" in js, "the viewer's key handler has no editing guard"
+
+
+@pytest.mark.parametrize("route", ["booth_answer", "booth_note", "booth_flag",
+                                   "booth_unmark", "booth_import_asks"])
+def test_mark_writes_do_not_block_the_event_loop(route):
+    """Every mark write takes a blocking `flock` and does synchronous disk I/O.
+    In an `async def` handler that runs ON the event loop, so a lock held by
+    another process — the CLI mid-`marks-import`, a second browser tab — freezes
+    every other request, including the index and `/healthz`.
+
+    Structural, like `test_stdlib_only`, and for the same reason: the failure is
+    a property of where the call runs, which no single-process response
+    assertion can see. The rule is that an async mark-write handler hands the
+    locked section to a worker thread and never calls the writer inline.
+    """
+    src = pathlib.Path(__file__).parent.parent / "booth" / "app.py"
+    fn = next(
+        n for n in ast.walk(ast.parse(src.read_text()))
+        if isinstance(n, ast.AsyncFunctionDef) and n.name == route
+    )
+    writers = {"answer_pick", "write_note", "set_flag", "delete_mark",
+               "import_legacy_asks"}
+    for node in ast.walk(fn):
+        if not isinstance(node, ast.Call):
+            continue
+        name = getattr(node.func, "id", None) or getattr(node.func, "attr", None)
+        if name in writers:
+            pytest.fail(f"{route} calls {name}() on the event loop; "
+                        "dispatch it through run_in_threadpool")
+
+
+def test_an_unreadable_mark_is_visible_on_the_page(client):
+    """Surviving the poisoned file is half of it. A note whose stored `text` is
+    unreadable hydrates with empty text, and the panel rendered that as an empty
+    `<pre>` with a withdraw button beside it — which looks exactly like a note
+    the operator wrote and then cleared.
+
+    `_hydrate`'s own docstring forbids this for picks ("a broken question the
+    session believes it posted has to be visible — silently hiding it is the one
+    outcome nobody can debug"). It is the same argument for every shape."""
+    c, data = client
+    b = data / "b"
+    b.mkdir()
+    (b / MARKS_FILE).write_text(json.dumps({
+        "version": 1,
+        "marks": [{"id": "n1", "shape": "note", "text": {"oops": True},
+                   "created": "2026-09-21T00:00:00+00:00"}],
+    }))
+
+    html = c.get("/b/b/").text
+    assert "⚠ broken" in html, "an unreadable mark rendered as an empty note"
+    assert "n1" in html
+
+
+def test_a_marks_file_no_one_can_parse_does_not_take_down_the_index(tmp_path):
+    """The v0.2.2 round adopted the RecursionError finding and closed only half
+    of it. `_hydrate_safe` guards hydration; `json.loads` runs BEFORE that, in
+    `_read_raw`, whose `except (OSError, ValueError, UnicodeDecodeError)` does
+    not cover RecursionError or MemoryError.
+
+    So a 400 KB file of nothing but brackets, in any one booth, still returned
+    500 for `/` and `/healthz` across every booth on the service. Found by the
+    U5 code-review panel against the sibling module and confirmed by running it.
+    The read is bounded now and both classes are caught.
+    """
+    booth = tmp_path / "b"
+    booth.mkdir()
+    (booth / MARKS_FILE).write_text("[" * 200_000 + "]" * 200_000)
+
+    assert marks_for(booth) == []
+
+
+def test_a_marks_file_too_large_to_be_marks_is_refused_before_it_is_read(tmp_path):
+    """Bounded by `stat`, not survived. A booth holds one marks document, and
+    the index reads every booth's on every page load."""
+    from booth.marks import MARKS_MAX_BYTES
+
+    booth = tmp_path / "b"
+    booth.mkdir()
+    (booth / MARKS_FILE).write_text(" " * (MARKS_MAX_BYTES + 10))
+
+    assert marks_for(booth) == []
+
+
+def test_a_write_over_an_unparseable_marks_file_still_refuses(tmp_path):
+    """The strict half of the asymmetry has to see the same failures the lenient
+    half does, or a file that reads as "no marks" gets replaced by a write that
+    believed it. Same two exception classes, same bound."""
+    from booth.marks import MarksCorrupt, set_flag
+
+    booth = tmp_path / "b"
+    booth.mkdir()
+    (booth / MARKS_FILE).write_text("[" * 200_000 + "]" * 200_000)
+
+    with pytest.raises(MarksCorrupt):
+        set_flag(booth, "a.png", True)
+
+
+# ---- findings from the U5 diff-scoped BUG-HUNT panel, 2026-09-22 ------------
+
+
+def test_the_marks_reader_never_blocks_on_a_file_that_is_not_a_file(tmp_path):
+    """Same hole the size cap opened in the manifest, in the sibling it was
+    copied from. `st_size` is 0 for a FIFO, so it passes the cap, and then
+    `read_text` blocks with no EOF. `list_booths` reads every booth's marks on
+    every `GET /` and `/healthz`."""
+    import os
+    import signal
+
+    booth = tmp_path / "b"
+    booth.mkdir()
+    os.mkfifo(booth / MARKS_FILE)
+
+    def _timeout(signum, frame):
+        raise AssertionError("marks_for blocked on a FIFO and never returned")
+
+    old = signal.signal(signal.SIGALRM, _timeout)
+    signal.alarm(5)
+    try:
+        assert marks_for(booth) == []
+    finally:
+        signal.alarm(0)
+        signal.signal(signal.SIGALRM, old)
+
+
+def test_new_marks_and_imported_marks_share_one_stamp_format(tmp_path):
+    """The v0.2.2 fix for the legacy-import ordering opened a NEW ordering bug,
+    which is the shape worth remembering. `import_legacy_asks` moved to
+    microsecond precision while `now_stamp` stayed at whole seconds, and `-` is
+    0x2D against `.` at 0x2E — so `...T10:00:00-07:00` sorts BEFORE
+    `...T10:00:00.500000-07:00`, putting a LATER mark ahead of an EARLIER
+    import inside the same second.
+
+    Deterministic order is a v1 invariant precisely because the operator refers
+    to things positionally. One format, or the rule cannot be stated.
+    """
+    from booth.marks import now_stamp
+
+    stamp = now_stamp()
+    assert "." in stamp.split("T")[1], f"now_stamp is not sub-second: {stamp}"
+    assert len(stamp.split(".")[1].split("+")[0].split("-")[0]) == 6
+
+
+def test_the_importer_cannot_raise_out_of_a_poisoned_entry(tmp_path):
+    """`marks_for` routes every entry through `_hydrate_safe`; the importer's
+    return still went through the bare `_hydrate`, so the one path that reads
+    entries it did not write was the one without the guard."""
+    booth = tmp_path / "b"
+    booth.mkdir()
+    (booth / MARKS_FILE).write_text(json.dumps({
+        "version": 1,
+        "marks": [{"id": "n1", "shape": "note", "text": {"bad": True},
+                   "created": "2026-09-21T00:00:00+00:00"}],
+    }))
+    (booth / f"q1{ASK_SUFFIX}").write_text(json.dumps(_single()))
+
+    from booth.marks import import_legacy_asks
+    out = import_legacy_asks(booth)          # must not raise
+    assert isinstance(out, list)
+
+
+def test_a_document_that_would_not_read_back_is_refused_at_the_write(tmp_path):
+    """The read bound is on the STORED bytes and the write adds `indent=2`, so a
+    document that fits in memory can land over the limit on disk and then read
+    back as no marks at all — every mark in the booth gone, silently. Refuse
+    loudly instead: a write that fails is recoverable.
+
+    Asserted against `_write_raw` directly, because no single mark can get
+    there: `_clean_text` caps a note at TEXT_MAX and a flag is a fixed shape.
+    The reachable path is accumulation — `_note_id` puts no ceiling on how many
+    notes one booth may carry — which is thousands of writes, not one. Testing
+    it through `write_note` would need a fixture nobody could justify, and
+    would be testing the cap rather than the guard.
+    """
+    from booth.marks import MARKS_MAX_BYTES, MarksCorrupt, _write_raw
+
+    booth = tmp_path / "b"
+    booth.mkdir()
+    bulk = [{"id": f"note-{i}", "shape": "note", "text": "x" * 500,
+             "created": "2026-09-21T00:00:00.000000+00:00"}
+            for i in range(MARKS_MAX_BYTES // 400)]
+
+    with pytest.raises(MarksCorrupt):
+        _write_raw(booth, bulk)
+    assert not (booth / MARKS_FILE).exists(), "a refused write still landed"
+
+
+def test_a_clock_restore_that_fails_does_not_take_the_route_down(tmp_path):
+    """The concrete half of the mtime-restore finding.
+
+    `_Locked.__enter__` puts the booth directory's clock back after creating its
+    lock, and `os.utime` can fail — a read-only directory, a booth whose owner
+    we are not. It used to escape into the route and answer 500 for what is
+    otherwise a perfectly good request. Not putting the clock back is a cost
+    this module can absorb; not answering is not.
+
+    The RACE half of that finding is documented in the code and deliberately not
+    closed: the alternative fix would silently retire the documented behaviour
+    that releasing a kept board resets its clock
+    (`test_releasing_a_board_RESETS_its_ttl_clock` pins that on purpose), which
+    is a TTL doctrine change rather than a bug fix.
+    """
+    import os
+
+    from booth.marks import MARKS_LOCK, set_flag
+
+    booth = tmp_path / "b"
+    booth.mkdir()
+    real_utime = os.utime
+
+    def boom(path, *a, **kw):
+        if str(path) == str(booth):
+            raise PermissionError("read-only directory")
+        return real_utime(path, *a, **kw)
+
+    os.utime = boom
+    try:
+        assert set_flag(booth, "a.png", True) is not None
+    finally:
+        os.utime = real_utime
+
+    assert (booth / MARKS_LOCK).exists()
+    assert [m.target for m in marks_for(booth)] == ["a.png"]
+
+
+# ---- the wrong-shaped answer, closed at the hydration boundary --------------
+
+
+def test_a_wrong_shaped_answer_is_an_error_at_hydration_not_a_500(tmp_path):
+    """A `.marks.json` that is well-formed JSON with a wrong-shaped `answer`
+    passed every reader and then raised in the TEMPLATE: `_hydrate` checked only
+    that `answer` was a dict, never that `answer["answers"]` was one, so
+    `marks_for` and `hold_read` both reported the mark healthy with no read
+    error — and `_ask_inline.html` asked a list for `.get`.
+
+    Measured at `42ea67f`, so it predates U3. U3 guarded its own surface with
+    `_safe_fragments` and left the gallery and marks pages alone by scope. This
+    closes it at the boundary the rest of the module already argues for: ONE
+    predicate, ONE place, every surface inherits it.
+
+    Defeating change: restoring the bare `isinstance(answer, dict)` check —
+    under which `error` is None here and both pages 500.
+    """
+    declare_pick(tmp_path, "batch", {"title": "T", "questions": [
+        {"key": "r1", "prompt": "A?", "options": ["x", "y"]},
+        {"key": "r2", "prompt": "B?", "options": ["x", "y"]}]})
+    raw = json.loads((tmp_path / ".marks.json").read_text())
+    for e in raw["marks"]:
+        if e["id"] == "batch":
+            e["answer"] = {"answers": [], "notes": ""}
+    (tmp_path / ".marks.json").write_text(json.dumps(raw))
+
+    mark = {m.id: m for m in marks_for(tmp_path)}["batch"]
+    assert mark.error, "a wrong-shaped answer hydrated as healthy"
+    assert "answer" in mark.error
+    # AND the mark is not silently emptied — the declaration survives, so the
+    # operator can still see WHICH question broke rather than a bare error.
+    assert mark.declaration is not None
+
+
+def test_a_healthy_multi_answer_still_hydrates(tmp_path):
+    """The other direction, so the guard cannot be satisfied by rejecting
+    everything. Defeating change: requiring `answers` unconditionally, which
+    would break every single-question pick."""
+    declare_pick(tmp_path, "multi", {"title": "T", "questions": [
+        {"key": "r1", "prompt": "A?", "options": ["x", "y"]}]})
+    declare_pick(tmp_path, "single", {"prompt": "Which?", "options": ["x", "y"]})
+    raw = json.loads((tmp_path / ".marks.json").read_text())
+    for e in raw["marks"]:
+        if e["id"] == "multi":
+            e["answer"] = {"answers": {"r1": {"choice": "x", "notes": ""}}, "notes": ""}
+        if e["id"] == "single":
+            e["answer"] = {"choice": "x", "notes": ""}
+    (tmp_path / ".marks.json").write_text(json.dumps(raw))
+    by = {m.id: m for m in marks_for(tmp_path)}
+    assert by["multi"].error is None, by["multi"].error
+    assert by["single"].error is None, by["single"].error
+
+
+def test_the_package_version_carries_no_literal_of_its_own():
+    """`booth.__version__` said `0.1.0` through six releases while pyproject
+    said `0.6.1` — a second copy of one fact, drifting silently, found only
+    while cutting 1.0.
+
+    THE ASSERTION IS THE ABSENCE OF A LITERAL, not agreement with pyproject:
+    `__version__` is now READ from pyproject, so comparing the two would be
+    circular and would prove only that the read works. The defeating change is
+    hardcoding a number back into this module, and that is what this catches.
+    """
+    src = (pathlib.Path(__file__).parent.parent / "booth" / "__init__.py").read_text()
+    literals = re.findall(r'__version__\s*=\s*["\']([^"\']+)["\']', src)
+    assert not literals, f"booth/__init__.py hardcodes a version again: {literals}"
+
+
+def test_the_package_version_is_the_one_the_tree_declares():
+    """And it resolves, from the tree, to what pyproject says — NOT to whatever
+    a stale dist-info in some venv happens to record. Found saying `0.3.0` from
+    installed metadata while the tree was at `1.0.0b1`."""
+    import booth
+
+    declared = tomllib.loads(
+        (pathlib.Path(__file__).parent.parent / "pyproject.toml").read_text()
+    )["project"]["version"]
+    assert booth.__version__ == declared
+    assert booth.__version__ != "0.0.0+unknown", "the pyproject read fell through"

tests/test_navigation.py

@@ -0,0 +1,404 @@
+"""U7 — the rail, the filters, the grid keyboard, and the groups.
+
+All four components. The fourth — replacing directory sections with
+filename-derived groups — was a scope DEPARTURE from ROADMAP's U7 row and was
+ratified by the operator on 2026-09-22; `test_no_group_rail_is_shipped_yet`,
+the guard that held it back while the ruling was outstanding, was deleted in
+the commit that built it. A guard that outlives its reason is worse than no
+guard, because the next reader trusts it.
+
+`unanswered` is taken to mean HAS AN OPEN PICK — the U4 hold predicate, which
+already exists and already has a home. The alternative reading ("has no mark at
+all") is a real and different question and is the contract's open question.
+"""
+
+from __future__ import annotations
+
+import json
+import pathlib
+import sys
+
+import pytest
+from fastapi.testclient import TestClient
+
+sys.path.insert(0, str(pathlib.Path(__file__).parent.parent))
+
+from booth.app import create_app  # noqa: E402
+from booth.marks import declare_pick, set_flag, write_note  # noqa: E402
+
+PNG = b"\x89PNG\r\n\x1a\n"
+
+
+@pytest.fixture
+def gallery(tmp_path):
+    """A booth with one of each: flagged, annotated, open pick, and plain."""
+    b = tmp_path / "g"
+    b.mkdir()
+    for n in ("a.png", "b.png", "c.png", "d.png"):
+        (b / n).write_bytes(PNG)
+    set_flag(b, "a.png", True)
+    write_note(b, "b.png", "a remark")
+    declare_pick(b, "q", {"prompt": "Which?", "options": ["x", "y"]}, target="c.png")
+    app = create_app(tmp_path, ttl_hours=24, start_sweeper=False)
+    return TestClient(app), b
+
+
+def _tiles(body: str) -> list[str]:
+    """The rels the grid actually rendered, in render order."""
+    import re
+    # `data-item` already exists on every tile (both the doc and media
+    # variants). Reusing it rather than adding a parallel `data-rel` is the
+    # same one-fact-one-place discipline INV-1 states for item facts.
+    return re.findall(r'data-item="([^"]+)"', body)
+
+
+def test_the_rail_counts_every_filter(gallery):
+    c, _ = gallery
+    body = c.get("/b/g/").text
+    assert 'class="rail"' in body
+    for token in ("all", "flagged", "annotated", "unanswered"):
+        assert f'data-filter="{token}"' in body, token
+
+
+@pytest.mark.parametrize("flt,expected", [
+    ("all", ["a.png", "b.png", "c.png", "d.png"]),
+    ("flagged", ["a.png"]),
+    ("annotated", ["b.png"]),
+    ("unanswered", ["c.png"]),
+])
+def test_a_filter_narrows_the_grid_server_side(gallery, flt, expected):
+    """INV-4: a filter is a LINK, not a script. Fetched directly, with no JS
+    executed, the server must return the narrowed grid.
+
+    Defeating change: binding filters to a click handler and returning the full
+    grid for every URL — under which this test gets four tiles every time."""
+    c, _ = gallery
+    assert _tiles(c.get(f"/b/g/?filter={flt}").text) == expected
+
+
+def test_filtering_never_reorders(gallery):
+    """INV-2, the load-bearing one. Grouping and filtering are VIEWS.
+
+    The defeating change is sorting the grid by anything derived from the
+    filter — which looks right and silently changes what "the third one" means,
+    the misfiled-judgment failure CLAUDE.md invariant 6 exists to prevent.
+
+    Asserted as a SUBSEQUENCE rather than a set: order is the property, so a
+    filter that returned the right tiles in the wrong sequence must go red."""
+    c, _ = gallery
+    # ⚠ THE BASELINE IS COMPUTED INDEPENDENTLY, and that is the whole test.
+    # The first version of this compared each filtered view against the
+    # UNFILTERED RESPONSE — and a mutation that reversed the order reversed
+    # both sides, so it stayed green under the exact change it forbade. Caught
+    # by running the mutation rather than trusting the assertion, which is the
+    # discipline in persistent-memory.d/2026-09-22-vacuous-falsifiers.md and
+    # which this test failed first time out.
+    #
+    # The independent truth is U1 INV-3: the item order IS `sorted(rel)`. So
+    # each filtered view must be sorted, full stop, with no reference to any
+    # other response.
+    for flt in ("all", "flagged", "annotated", "unanswered"):
+        got = _tiles(c.get(f"/b/g/?filter={flt}").text)
+        assert got == sorted(got), f"{flt} rendered out of sorted(rel) order: {got}"
+    # and every filtered view is a SUBSEQUENCE of the true order, not a reshuffle
+    every = sorted(["a.png", "b.png", "c.png", "d.png"])
+    for flt in ("all", "flagged", "annotated", "unanswered"):
+        got = _tiles(c.get(f"/b/g/?filter={flt}").text)
+        assert got == [r for r in every if r in got], flt
+
+
+def test_an_unknown_filter_falls_back_to_all_and_does_not_500(gallery):
+    """A filter arrives from a URL, which is operator-editable and link-shared.
+    Defeating change: indexing a dict by the raw parameter."""
+    c, _ = gallery
+    for junk in ("nonsense", "", "../../etc", "flagged;drop"):
+        r = c.get(f"/b/g/?filter={junk}")
+        assert r.status_code == 200, junk
+        assert len(_tiles(r.text)) == 4, junk
+
+
+def test_the_zoom_ring_is_identical_under_every_filter(gallery):
+    """The ring is the item order filtered to images and must not notice the
+    grid's filter — otherwise `next` means something different depending on how
+    the operator arrived, and a flag lands on the wrong artifact.
+
+    Defeating change: building the ring from the filtered list."""
+    c, _ = gallery
+    rings = set()
+    for flt in ("all", "flagged", "annotated", "unanswered"):
+        c.get(f"/b/g/?filter={flt}")
+        body = c.get("/b/g/b.png?view=1").text
+        import re
+        rings.add(tuple(re.findall(r'href="([^"]*\.png[^"]*)"', body)))
+    assert len(rings) == 1, f"the ring changed with the filter: {rings}"
+
+
+def test_the_rail_is_absent_on_a_booth_with_no_grid(tmp_path):
+    """INV-5's sibling: a rail over nothing is chrome. The standing link board
+    has no items, so it must not render one."""
+    b = tmp_path / "links"
+    b.mkdir()
+    (b / "links.md").write_text("- [r](https://x.test/) <sub>· a · 2026-09-01 00:00</sub>\n")
+    c = TestClient(create_app(tmp_path, ttl_hours=24, start_sweeper=False))
+    assert 'class="rail"' not in c.get("/b/links/").text
+
+
+def test_the_keyboard_is_not_bound_when_there_is_no_grid(tmp_path):
+    """INV-5. Defeating change: binding the handler unconditionally, so `f` on
+    the standing link board swallows the keystroke and flags nothing."""
+    b = tmp_path / "links"
+    b.mkdir()
+    (b / "links.md").write_text("- [r](https://x.test/) <sub>· a · 2026-09-01 00:00</sub>\n")
+    c = TestClient(create_app(tmp_path, ttl_hours=24, start_sweeper=False))
+    assert "gridkeys" not in c.get("/b/links/").text
+
+
+def test_the_keyboard_is_bound_when_there_is_one(gallery):
+    c, _ = gallery
+    assert "gridkeys" in c.get("/b/g/").text
+
+
+
+
+# --- U7 slice 2: the groups ----------------------------------------------
+
+
+@pytest.fixture
+def grouped(tmp_path):
+    """Two groups whose members INTERLEAVE in `sorted(rel)`.
+
+    `a/x1.png, a/y1.png, b/x2.png, b/y2.png` is the sorted order; group `x` is
+    at positions 0 and 2, group `y` at 1 and 3. That interleaving is the whole
+    point of the fixture — a grid re-sorted by `(group, rel)` to make groups
+    render contiguously would pass every set-based assertion and fail these.
+    """
+    b = tmp_path / "g"
+    for rel in ("a/x1.png", "a/y1.png", "b/x2.png", "b/y2.png"):
+        p = b / rel
+        p.parent.mkdir(parents=True, exist_ok=True)
+        p.write_bytes(PNG)
+    app = create_app(tmp_path, ttl_hours=24, start_sweeper=False)
+    return TestClient(app), b
+
+
+def _groups(body: str) -> list[str]:
+    """The group keys the rail listed, in render order."""
+    import re
+    return re.findall(r'data-group="([^"]+)"', body)
+
+
+def test_the_rail_lists_groups_when_grouping_is_informative(grouped):
+    c, _ = grouped
+    body = c.get("/b/g/").text
+    assert 'class="rail-groups"' in body
+    assert _groups(body) == ["x", "y"]
+
+
+def test_group_order_is_the_position_of_the_first_member(tmp_path):
+    """The settled rule (ROADMAP, operator 2026-09-22): groups order by where
+    each group's FIRST member falls in the rendered sequence.
+
+    ⚠ THIS FIXTURE IS BUILT SO THE THREE PLAUSIBLE RULES ALL DISAGREE. The
+    first version used `w, x, y` — whose positional order happens to BE
+    alphabetical, so it stayed green under the very change it forbade. Caught
+    by running the mutation, not by reading the assertion; the same trap
+    persistent-memory.d/2026-09-22-vacuous-falsifiers.md names and the same one
+    `test_filtering_never_reorders` fell into an hour after it was written.
+
+        sorted(rel):  a/z1  a/z2  b/a1  b/a2  b/a3  c/m1  c/m2
+        by position:  z (0), a (2), m (5)      <- the rule
+        alphabetical: a, m, z                  <- wrong, and differs
+        by count:     a(3), z(2), m(2)         <- wrong, and differs
+    """
+    b = tmp_path / "g"
+    for rel in ("a/z1.png", "a/z2.png", "b/a1.png", "b/a2.png", "b/a3.png",
+                "c/m1.png", "c/m2.png"):
+        q = b / rel
+        q.parent.mkdir(parents=True, exist_ok=True)
+        q.write_bytes(PNG)
+    c = TestClient(create_app(tmp_path, ttl_hours=24, start_sweeper=False))
+    assert _groups(c.get("/b/g/").text) == ["z", "a", "m"]
+
+
+def test_grouping_never_reorders_the_grid(grouped):
+    """INV-2, the load-bearing one.
+
+    The defeating change is sorting the grid by `(group, rel)` so groups render
+    contiguously — which looks right, passes any set comparison, and silently
+    changes what "the third one" means. This fixture interleaves precisely so
+    that change goes red.
+
+    The baseline is INDEPENDENT (U1 INV-3: the order IS `sorted(rel)`), not a
+    second response — the vacuous-falsifier trap this suite already fell into
+    once."""
+    c, _ = grouped
+    tiles = _tiles(c.get("/b/g/").text)
+    assert tiles == ["a/x1.png", "a/y1.png", "b/x2.png", "b/y2.png"]
+    assert tiles == sorted(tiles)
+
+
+def test_every_group_anchor_lands_on_a_rendered_tile(grouped):
+    """A jump-to-group link that scrolls nowhere is worse than no link. Every
+    anchor must name an id the page actually carries.
+
+    Defeating change: anchoring to the group KEY (`#group-x`) while the tiles
+    carry `id="item-<rel>"` — which renders, looks right, and does nothing."""
+    import re
+    c, _ = grouped
+    body = c.get("/b/g/").text
+    hrefs = re.findall(r'class="rail-g"[^>]*href="#([^"]+)"', body)
+    assert hrefs, "the rail rendered no group anchors"
+    for h in hrefs:
+        assert f'id="{h}"' in body, f"anchor #{h} names no element on the page"
+
+
+def test_no_group_rail_when_every_item_is_its_own_group(gallery):
+    """INV-3's real failure mode, and it is NOT the one the contract feared.
+
+    `a.png b.png c.png d.png` yields four groups of one — a rail that is a
+    second copy of the grid. Measured live: `pewpew-ui-brief` gives 23 groups
+    for 34 items, `dfa-concepts` 13 for 20. The contract only guarded the
+    opposite degeneracy (one group for everything), which is why this test
+    exists.
+
+    Defeating change: `{% if rail.groups %}`, true for four singletons."""
+    c, _ = gallery
+    body = c.get("/b/g/").text
+    assert 'class="rail-groups"' not in body
+    assert 'class="rail"' in body, "the filter rail must still be here"
+
+
+def test_no_group_rail_when_there_is_only_one_group(tmp_path):
+    """INV-3 as the contract states it, with the live specimen: `sc-iso-spread`
+    is `DSC0001.jpg` through `DSC0006.jpg` — one group, six images.
+
+    Defeating change: `{% if rail.groups %}`, true for a single group."""
+    b = tmp_path / "flat"
+    b.mkdir()
+    for i in range(1, 7):
+        (b / f"DSC{i:04d}.jpg").write_bytes(PNG)
+    c = TestClient(create_app(tmp_path, ttl_hours=24, start_sweeper=False))
+    body = c.get("/b/flat/").text
+    assert 'class="rail-groups"' not in body
+    assert 'class="rail"' in body
+
+
+def test_groups_describe_the_filtered_grid(tmp_path):
+    """The rail describes what is ON SCREEN. An anchor to a group the filter
+    has hidden would scroll nowhere — the same defect as a wrong id, arriving
+    by a different route.
+
+    Three groups of two; the flag covers `x` and `y` entirely and `z` not at
+    all. Under `?filter=flagged` the rail must list x and y and MUST NOT list
+    z, whose two tiles are not on the page.
+
+    Defeating change: deriving groups from the full gallery rather than from
+    the rendered list — under which `z` appears and its anchor goes nowhere."""
+    b = tmp_path / "g"
+    b.mkdir()
+    for n in ("x1.png", "x2.png", "y1.png", "y2.png", "z1.png", "z2.png"):
+        (b / n).write_bytes(PNG)
+    for n in ("x1.png", "x2.png", "y1.png", "y2.png"):
+        set_flag(b, n, True)
+    c = TestClient(create_app(tmp_path, ttl_hours=24, start_sweeper=False))
+
+    assert _groups(c.get("/b/g/").text) == ["x", "y", "z"]
+    body = c.get("/b/g/?filter=flagged").text
+    assert _groups(body) == ["x", "y"]
+    assert _tiles(body) == ["x1.png", "x2.png", "y1.png", "y2.png"]
+
+
+def test_a_filtered_view_too_small_to_group_drops_the_group_row(grouped):
+    """The informativeness rule binds to the RENDERED list, not to the booth.
+
+    One flagged tile is one group of one, which cannot navigate — so the group
+    row goes away even though the unfiltered booth has a perfectly good one.
+    The filter rail stays, because that is how the operator gets back."""
+    c, b = grouped
+    set_flag(b, "a/x1.png", True)
+    assert 'class="rail-groups"' in c.get("/b/g/").text
+    body = c.get("/b/g/?filter=flagged").text
+    assert 'class="rail-groups"' not in body
+    assert 'class="rail"' in body
+
+
+def test_the_zoom_ring_ignores_grouping(grouped):
+    """The ring is `sorted(rel)` filtered to images and must not notice groups
+    any more than it notices filters.
+
+    THE FIXTURE IS THE FALSIFIER. From `a/x1.png`, sorted order says next is
+    `a/y1.png` — a DIFFERENT group. A ring rebuilt per group would say
+    `b/x2.png`, the next member of group `x`, and `→` would start walking a
+    sequence the operator never saw on the page. That is invariant 6's
+    misfiled-judgment failure exactly: the flag lands on the wrong artifact."""
+    import re
+    c, _ = grouped
+    body = c.get("/b/g/view?f=a/x1.png").text
+    nxt = re.findall(r'class="vnav vnext" href="\?f=([^"&]+)"', body)
+    assert nxt == ["a/y1.png"], f"the ring followed the group, not sorted(rel): {nxt}"
+    # and the zoom page has no group chrome at all — it is one artifact, not a wall
+    assert "data-group" not in body
+
+
+def test_no_route_body_derives_a_group(gallery):
+    """INV-1, the same assertion U1 makes for `classify` and `render_doc`.
+
+    Defeating change: a route or template computing a prefix inline — the
+    caption bug in a new field."""
+    import inspect
+
+    import booth.app as app_mod
+
+    src = inspect.getsource(app_mod.create_app)
+    assert "_group_of" not in src, "create_app must read Item.group, not derive it"
+
+
+def test_a_hostile_filename_cannot_break_out_of_the_rail(tmp_path):
+    """Group keys and anchors are AGENT-AUTHORED — they are filenames, and a
+    session makes a booth by making a folder with no validation anywhere in the
+    path. CLAUDE.md names autoescape as load-bearing for exactly this.
+
+    Defeating change: building the rail markup with `|safe`, or assembling the
+    href by string concatenation outside Jinja. Both render, both look right,
+    and both put attacker-controlled bytes into an attribute."""
+    b = tmp_path / "g"
+    b.mkdir()
+    for n in ('q"x1.png', 'q"x2.png', "s<script>1.png", "s<script>2.png"):
+        (b / n).write_bytes(PNG)
+    c = TestClient(create_app(tmp_path, ttl_hours=24, start_sweeper=False))
+    import re
+    r = c.get("/b/g/")
+    assert r.status_code == 200
+    body = r.text
+
+    # THE RAIL ITSELF, isolated — asserting over the whole page would pass on a
+    # booth where the escaping happened somewhere else.
+    nav = re.search(r'<nav class="rail-groups".*?</nav>', body, re.S)
+    assert nav, "the rail rendered no group row"
+    nav = nav.group(0)
+
+    # No tag the template did not write, and no attribute the filename closed.
+    # Asserted as the SET of element names rather than by counting `<`, which
+    # the first version got wrong by forgetting the `<b>` counts — an arithmetic
+    # slip that made the test red for a reason unrelated to escaping.
+    tags = set(re.findall(r"</?([a-zA-Z][a-zA-Z0-9]*)", nav))
+    assert tags == {"nav", "a", "b"}, f"the rail grew an element: {tags}"
+    assert 'data-group="q"' not in nav, "the quote closed the attribute"
+    assert "&lt;script&gt;" in nav and "<script" not in nav
+    assert "&#34;" in nav or "&quot;" in nav, "the quote was not escaped"
+
+
+def test_a_group_key_is_never_the_empty_string(tmp_path):
+    """`_group_of` returns None rather than "" for a stem with nothing before
+    the digits. A "" key would render a nameless rail row that files every
+    numbered render under it — the failure the None is there to prevent.
+
+    Defeating change: `return segs[0]` without the `or None`."""
+    b = tmp_path / "g"
+    b.mkdir()
+    for n in ("01.png", "02.png", "03.png", "ac1.png", "ac2.png"):
+        (b / n).write_bytes(PNG)
+    c = TestClient(create_app(tmp_path, ttl_hours=24, start_sweeper=False))
+    body = c.get("/b/g/").text
+    assert 'data-group=""' not in body
+    assert "" not in _groups(body)