fix: four defects the U4 bug-hunt panel found in code it did not add

All four pre-date U4 and sit in files it touched, which is why a diff-scoped
robustness lens saw them. They are separated from the unit's own commit so the
feature history stays readable; the release tags both.

* A booth name reached a JS string context. The confirm dialogs interpolated
  the name into a string literal inside `onsubmit`. Jinja's autoescape is
  HTML-attribute escaping, not JS-string escaping: the browser decodes the
  entity back to a quote before the JS parser sees it, so a name crafted to
  close the string executed on submit. Booth names are agent-authored — making
  a folder under the data dir is the whole API — so this was a live path, not a
  theoretical one. The name now travels as a data attribute to a delegated
  handler, where escaping is escaping.

* An unreadable `links.md` returned 500 for the whole booth page. `is_file()`
  then an unguarded `read_text()`. The board is one tile on that page, and a
  page that will not load is worse than one missing a tile — the posture
  `read_blurred`, `marks_for` and `read_manifest` already take.

* The index order had no tie-breaker, which violates the deterministic-order
  invariant. Equal-mtime booths fell back to whatever `iterdir()` yielded, and
  two booths landed by one `rsync` batch share an mtime exactly. Now
  `(mtime, name)` reverse: newest first, then name. The operator refers to
  cards positionally, so a sequence that moves between renders misfiles his
  judgment rather than crashing.

* `/b/<n>/marks.json` reported damage as empty success. `booth marks` exits 3
  on an unreadable file precisely so a caller can tell "not yet" from "broken";
  the HTTP mirror — the only reader a remote session has — returned the same
  empty list for both. It now carries `error` and `detail`. The status stays
  200 deliberately: reads are lenient here, and a pinned status code is a
  promise to remote clients this fix has no business breaking.

Each has a regression test. 410 tests.

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.

README.md

@@ -15,7 +15,9 @@ 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
 
@@ -159,7 +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
@@ -168,6 +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)
 
@@ -438,11 +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 +1,7 @@
 # The Booth — roadmap
 
 Design: [`docs/design/information-architecture.md`](docs/design/information-architecture.md).
-Current version: `0.3.0` (U1, U2 and U5 landed; extracted from eshpfi 2026-09-21).
+Current version: `0.4.0` (U1, U2, U4 and U5 landed; extracted from eshpfi 2026-09-21).
 
 ## v1 target
 
@@ -13,7 +13,7 @@ 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 |
-| 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**~~ — **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 |
 | 7 | **Navigation at 270 items** — sections, rail, filters, grid keyboard | one flat wall; subfolder structure discarded at render | U7 |
@@ -22,7 +22,7 @@ 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, U2 and U5 are landed.** U3 and U4 are unblocked and unstarted; U6 remains
+**U1, U2, U4 and U5 are landed.** U3 is unblocked and unstarted; U6 remains
 independent and unstarted; U7 waits on the rest.
 
 **U5's adoption is a measured prediction, not a finished result**, and it is
@@ -74,6 +74,12 @@ Where it already binds, and what the rule is in each case:
 | link board rows | pinned first, then newest-first |
 | a booth's announcement | not a collection — one flat record per booth, nothing to order (U5) |
 
+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.
+
 Where it is still to be decided, and must be before the unit ships: **U7's
 section ordering and its compare pairing** (sections need a stated order among
 themselves, not just within; pairing by filename needs a rule for what happens

booth/app.py

@@ -7,15 +7,26 @@ Model (deliberately dead-simple, no database):
   * GET /b/<name>/   -> if <name>/index.html exists, serve it verbatim; otherwise
                         auto-render a gallery of the images / webm-videos / audio in it.
   * GET /b/<name>/<file> -> serve a file out of the booth (also feeds a custom index.html's assets).
-  * 24h TTL: a background sweeper wipes any booth untouched for TTL hours. A booth's
+  * LIFETIME IS DERIVED, not set by a boolean (U4). Three states, and `sweep_once`
-    age is measured from the *newest* mtime in its tree, so it lives while it's being
+    is the only thing that honours the first two:
-    worked on and self-destructs TTL hours after the last activity.
+      KEPT      `.forever` present. Never swept, own lane at the top of the index.
-  * KEPT BOOTHS: a booth containing the KEEP_MARKER dotfile (`.forever`) is exempt
+                Durable operator-facing boards — chiefly the standing link board,
-    from the sweep and renders in its own lane above the ephemeral grid. That is the
+                whose whole purpose is to outlive the scrollback it replaces.
-    home for durable operator-facing boards — chiefly the standing link board agent
+      HELD      an open pick in `.marks.json`, or marks that cannot be read at all.
-    sessions post to, whose whole purpose is to survive longer than the scrollback
+                A booth the operator still owes an answer to is not the sweeper's
-    it replaces. Opt-in per booth, so the ephemeral default is unchanged and nobody
+                to take, and one whose judgment we failed to READ is certainly not.
-    inherits a cleanup chore; `rm` the sentinel and the booth rejoins the sweep.
+      EPHEMERAL everything else: wiped TTL hours after the last activity. Age is the
+                *newest* mtime in the tree, so a booth lives while it is being
+                worked on and self-destructs once it stops.
+  * VIEWING IS ACTIVITY. A deliberate GET of a booth's own page writes VIEW_MARKER,
+    which the age rule already counts — if the operator is still looking at it, it
+    is still alive. Browsing the index is not a view, and neither is a session
+    polling `marks.json`: an agent must not be able to hold its own booth open.
+  * WHY DERIVED. `.forever` was the ONLY way to say three different things, and the
+    measurement showed it carrying all of them — 17 of 24 live booths on 2026-09-22
+    (70%, up from 54%), with three of the four booths awaiting an answer ALSO pinned
+    by hand. Only "this is durable" is what keep means. The other two are facts the
+    service already held and did not consult.
 
 State is the filesystem — `ls ~/booth-data` tells you everything. That is the whole point.
 """
@@ -34,6 +45,7 @@ import time
 import zipfile
 from contextlib import asynccontextmanager
 from pathlib import Path
+from typing import Sequence
 from urllib.parse import quote, unquote
 
 from fastapi import FastAPI, File, Form, HTTPException, Request, UploadFile
@@ -85,6 +97,14 @@ from booth.items import (  # noqa: E402,F401
 # flag to remember, no state anywhere but the filesystem.
 KEEP_MARKER = ".forever"
 
+# Records the last deliberate look at a booth (U4). A dotfile for the same two
+# reasons KEEP_MARKER is one — `booth_items` and `zip_booth` skip it, so it
+# costs nothing in counts, galleries or zips — and NOT a `.lock` dotfile, so
+# `_newest_mtime` COUNTS it and the existing age rule picks the view up with no
+# new arithmetic. That is the whole integration: a view is one more thing in
+# the tree, not a second term in the formula.
+VIEW_MARKER = ".viewed"
+
 # ⚠⚠ BLUR IS COSMETIC, NOT ACCESS CONTROL. The file is still served at its own
 # URL, still in the zip, still on disk. This hides an item from a glance — a
 # shoulder, a screen-share, a scroll past something you did not want to see
@@ -124,12 +144,14 @@ from booth.asks import (  # noqa: E402
 )
 from booth.marks import (  # noqa: E402
     MARKS_FILE,
+    Mark,
     MarksCorrupt,
     answer_pick,
     as_dict,
     declare_pick,
     delete_mark,
     import_legacy_asks,
+    hold_read,
     marks_for,
     marks_for_target,
     open_marks,
@@ -189,18 +211,34 @@ def _newest_mtime(path: Path) -> float:
     gets its clock counted. Everything else counts too, dotfiles included,
     because `.marks.json`, `.blurred` and `.pins` are the operator doing
     something.
+
+    ⚠ A STAT WE CANNOT DO READS AS *FRESH*, NEVER AS EPOCH-OLD. This function
+    feeds `is_expired`, which feeds `rmtree`. Returning 0.0 for a booth whose
+    own stat fails made it maximally ancient and therefore the FIRST thing the
+    sweeper takes — a permissions or ELOOP problem resolving to a deletion. The
+    bug-hunt panel found this as one of four paths into the same shape. Not
+    knowing a booth's age is a reason to leave it alone.
+
+    `FileNotFoundError` on an entry is the exception, and it stays a skip: a
+    dangling symlink and a file removed mid-scan both raise it, and neither is
+    a thing with an mtime worth counting. Any OTHER per-entry OSError means we
+    could not read something that IS there, so the age is unknowable and the
+    booth reads as fresh.
     """
+    now = time.time()
     try:
         newest = path.stat().st_mtime
     except OSError:
-        return 0.0
+        return now
     for p in path.rglob("*"):
         if p.name.startswith(".") and p.name.endswith(".lock"):
             continue
         try:
             m = p.stat().st_mtime
+        except FileNotFoundError:
+            continue                      # dangling symlink, or gone mid-scan
         except OSError:
-            continue
+            return now                    # cannot read it — cannot judge the age
         if m > newest:
             newest = m
     return newest
@@ -222,8 +260,102 @@ def is_expired(path: Path, ttl_seconds: float, now: float | None = None) -> bool
 
 
 def is_kept(path: Path) -> bool:
-    """True if this booth carries the keep sentinel and must never be swept."""
+    """True if this booth carries the keep sentinel and must never be swept.
-    return (path / KEEP_MARKER).exists()
+
+    `lstat`, not `Path.exists()`, and an unreadable answer counts as KEPT. The
+    old form collapsed ELOOP and EACCES into False, so a kept booth whose
+    sentinel could not be stat'd became eligible for the sweep — a failed read
+    authorizing a delete, which is the shape the bug-hunt panel found four ways
+    into. `lstat` also means a `.forever` SYMLINK counts, dangling or not:
+    somebody put it there to mean keep.
+    """
+    try:
+        (path / KEEP_MARKER).lstat()
+        return True
+    except FileNotFoundError:
+        return False
+    except OSError:
+        return True
+
+
+def record_view(booth: Path) -> None:
+    """Note that somebody deliberately looked at this booth (U4).
+
+    Touches VIEW_MARKER and lets `_newest_mtime` do the rest — a view enters
+    the age rule as a file in the tree, not as a new term in the arithmetic.
+
+    NEVER RAISES. A read-only mount, a booth owned by another uid, a full disk,
+    a booth deleted between the route's resolve and this call: every one of
+    those costs the timestamp, not the page. The same trade `_Locked.__enter__`
+    makes on its `os.utime`, and for the same stated reason — not recording the
+    look is a cost this service can absorb, not answering the request is not.
+    A booth whose view cannot be recorded simply ages on its content mtime,
+    which is what every booth did before this existed.
+    """
+    # O_NOFOLLOW, not `Path.touch()`. `touch` on an existing symlink follows it,
+    # so 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. Three of four bug-hunt arms found it independently. A symlink
+    # here now raises ELOOP into the swallow below: view-recording quietly stops
+    # for that booth, which is the right way to lose this argument.
+    #
+    # O_CREAT alone does not move the mtime of a file that already exists, so
+    # the utime is not decoration: the marker must read as NOW or the whole
+    # mechanism is a file nobody's clock looks at.
+    try:
+        fd = os.open(booth / VIEW_MARKER,
+                     os.O_WRONLY | os.O_CREAT | os.O_NOFOLLOW, 0o644)
+        try:
+            os.utime(fd)
+        finally:
+            os.close(fd)
+    except OSError:
+        pass
+
+
+HOLD_UNREADABLE = "unreadable"
+HOLD_OPEN = "open"
+
+
+def hold_reason(marks: Sequence[Mark], error: str | None) -> str | None:
+    """WHY this booth must not be swept, or None if it may be. THE hold predicate.
+
+    Returns a reason rather than a bool so the surface that has to say why can
+    read it off the same value the sweeper acts on. A boolean plus a separate
+    error string is two representations of one state, and they drift.
+
+    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 BOTH LEVELS OF DAMAGE, which is the correction the bug-hunt
+    panel forced (2026-09-22). `marks_for` is lenient because a review page that
+    will not load is worse than one missing an annotation — the right trade for
+    a RENDER and the wrong one for a DELETE, where the same leniency wipes the
+    booth whose judgment we had just failed to read, artifacts and all. The
+    first cut of this caught FILE-level damage only:
+
+      * file-level — `.marks.json` will not parse at all. `hold_read` reports it.
+      * ENTRY-level — the document parses, but one mark fails normalization and
+        `_hydrate_safe` hands back a `Mark` carrying `error`. `_is_open` returns
+        False for an errored pick, ON PURPOSE (a broken pick can never be
+        answered; the CLI spells that exit code 4) — so such a booth read as
+        `not held` and SWEPT, while the panel beside it rendered the broken mark
+        in full. Four arms found four ways into that shape; this was the worst.
+
+    A mark that cannot be read is judgment we cannot see. Deleting the booth it
+    belongs to is the one thing we must not do with it.
+
+    Openness itself is `open_marks` and nothing else (U2 INV-2): a partially
+    answered pick is STILL open and still holds, which is the reading that
+    makes this rule correct rather than one that sweeps a review in flight.
+    """
+    if error is not None or any(m.error is not None for m in marks):
+        return HOLD_UNREADABLE
+    if open_marks(marks):
+        return HOLD_OPEN
+    return None
 
 
 def sweep_once(data_dir: Path, ttl_seconds: float, now: float | None = None) -> list[str]:
@@ -232,10 +364,25 @@ def sweep_once(data_dir: Path, ttl_seconds: float, now: float | None = None) ->
     Only ever removes direct children of data_dir (never data_dir itself), and
     skips dotfolders so a stray control dir can opt out.
 
+    TWO exemptions, and this is the only function that honours either.
+
     A booth carrying KEEP_MARKER is exempt no matter how stale it is. That is
-    the one escape hatch from the 24h contract, and it is opt-in per booth: the
+    the explicit escape hatch, opt-in per booth: the default stays ephemeral, so
-    default stays ephemeral, so nobody inherits a cleanup chore they did not ask
+    nobody inherits a cleanup chore they did not ask for. Removing the sentinel
-    for. Removing the sentinel hands the booth straight back to the sweeper.
+    hands the booth straight back to the sweeper.
+
+    A booth that is HELD — an open pick, or marks we cannot read — is exempt for
+    as long as that holds (U4). This is the derived half: the operator was
+    pressing `.forever` to mean "not yet" because nothing else could say it, and
+    the service already knew. 17 of 24 live booths carried the sentinel on
+    2026-09-22, and three of the four booths in the fleet awaiting an answer
+    carried it too — the "not yet" case, caught in the act.
+
+    Reading the marks costs ONE strict read per booth per tick — `hold_read`,
+    which answers both halves of the hold question at once. It is deliberately
+    not two calls: two reads of one file are not one read of one state, and the
+    pair that loses that race is the pair that deletes. Do not "optimize" this
+    back into `marks_for` plus `read_error`.
     """
     wiped: list[str] = []
     if not data_dir.is_dir():
@@ -246,6 +393,8 @@ def sweep_once(data_dir: Path, ttl_seconds: float, now: float | None = None) ->
         try:
             if is_kept(child):
                 continue
+            if hold_reason(*hold_read(child)):   # ONE read — see hold_read
+                continue
             if is_expired(child, ttl_seconds, now):
                 shutil.rmtree(child)
                 wiped.append(child.name)
@@ -278,7 +427,20 @@ def list_booths(data_dir: Path, ttl_seconds: float, now: float | None = None) ->
         # flag a booth that is waiting on the operator. ONE file read per booth
         # — which is why marks live in one file per booth rather than a sidecar
         # per mark. This loop runs on every index page load.
-        marks = marks_for(child)
+        # ONE read for BOTH the badge and the lifetime decision. It has to be
+        # one: `marks_for` is lenient, so an unreadable `.marks.json` reads as
+        # no marks — fine for a card, wrong for the reaper, which would then
+        # delete the booth whose judgment it had just failed to read. And
+        # asking the two questions with two reads is not one read of one state:
+        # a write landing between them yields `([], None)`, the pair that
+        # deletes. `hold_read` answers both from one read; the lenient reader
+        # comes back only on the error path, where leniency is the point.
+        held_marks, read_err = hold_read(child)
+        # The DECISION comes from that one read and nothing else. The lenient
+        # re-read below is for DISPLAY only — feeding it back into the predicate
+        # would rebuild the two-read seam this call exists to close.
+        hold = hold_reason(held_marks, read_err)
+        marks = held_marks if read_err is None else marks_for(child)
         # The booth's own announcement — who posted it and why. One more small
         # read per booth, beside the marks read already here, and `read_manifest`
         # cannot raise for the same reason `marks_for` must not: this loop runs
@@ -314,11 +476,23 @@ def list_booths(data_dir: Path, ttl_seconds: float, now: float | None = None) ->
                 # tested `answer is None`, so a half-answered pick read as closed
                 # here while the panel beside it rendered `◐ partial`.
                 "marks_open": len(open_marks(marks)),
+                # U4: WHY this booth is or is not counting down. The card must
+                # never just stop the clock silently — `.forever` was at least
+                # visible as a lane, and an invisible rule would be worse than
+                # the boolean it replaces.
+                # WHY it is or is not counting down — the reason, not a bool
+                # beside a string that can disagree with it.
+                "hold": hold,
                 "expires_in": max(0.0, ttl_seconds - (now - mtime)),
                 "mtime": mtime,
             }
         )
-    booths.sort(key=lambda b: b["mtime"], reverse=True)
+    # Newest first, NAME as the tie-break. Sorting on mtime alone left equal-mtime
+    # booths ordered by whatever `iterdir()` yielded, which is not a rule — and
+    # invariant 6 is not "usually stable", it is a sentence you can write down.
+    # Two booths created by one `rsync` batch share an mtime exactly, and the
+    # operator refers to cards positionally.
+    booths.sort(key=lambda b: (b["mtime"], b["name"]), reverse=True)
     return booths
 
 
@@ -689,6 +863,10 @@ def create_app(
     @app.get("/b/{name}/", response_class=HTMLResponse)
     def booth_view(request: Request, name: str, download: int = 0):
         booth = resolve_booth(name)
+        # U4: viewing is activity. ABOVE both early returns — the zip download
+        # and the verbatim-index.html branch are looks at this booth too, and a
+        # verbatim report is the shape the operator stares at longest.
+        record_view(booth)
         if download:
             # whole-booth zip — the download path for a verbatim index.html booth
             # (which has no gallery/per-file chrome), and a "download all" for any.
@@ -721,7 +899,9 @@ def create_app(
             it for it in build_gallery(booth)
             if not ((booth / LINKS_FILE).is_file() and it["name"] == LINKS_FILE)
         ]
-        marks = marks_for(booth)
+        held_marks, read_err = hold_read(booth)  # ONE read; see list_booths
+        hold = hold_reason(held_marks, read_err)
+        marks = held_marks if read_err is None else marks_for(booth)
         return templates.TemplateResponse(
             request,
             "booth.html",
@@ -739,13 +919,13 @@ def create_app(
                 # Ordered pinned-first then newest-first, each row stamped with a
                 # `pinned` flag. Empty list for every other booth, so the template
                 # branch simply does not fire.
-                "board": (
+                # `is_file()` then an UNGUARDED read was a 500 waiting on a
-                    order_for_display(
+                # mode change or an EIO: the board is one tile on this page, and
-                        parse_link_entries((booth / LINKS_FILE).read_text()),
+                # a page that will not load is worse than one missing a tile —
-                        read_pins(booth),
+                # the same posture `read_blurred`, `marks_for` and
-                    )
+                # `read_manifest` already take. A booth whose `links.md` cannot
-                    if (booth / LINKS_FILE).is_file() else []
+                # be read renders as a booth with no board.
-                ),
+                "board": _board_rows(booth),
                 # Marks: operator judgment attached to this booth or to one of
                 # its items — a session's question (`pick`), the operator's own
                 # remark (`note`), the operator's selection (`flag`). Rendered
@@ -764,10 +944,29 @@ def create_app(
                 # a booth URL handed to the operator lands HERE, never on the
                 # index, and job 5 is "operator, look at this".
                 "manifest": read_manifest(booth),
+                # The lifetime line, same three states as the index card: a
+                # booth URL handed to the operator lands HERE, not on the index,
+                # so "why is this not counting down" has to be answerable here.
+                "hold": hold,
                 "expires_in": max(0.0, ttl_seconds - booth_age_seconds(booth)),
             },
         )
 
+    def _board_rows(booth: Path) -> list[dict]:
+        """The link board's rows, or [] for a board that cannot be read.
+
+        NEVER RAISES, for the reason every other read on this page does not:
+        one damaged file must cost its own tile, not the booth page."""
+        try:
+            if not (booth / LINKS_FILE).is_file():
+                return []
+            return order_for_display(
+                parse_link_entries((booth / LINKS_FILE).read_text()),
+                read_pins(booth),
+            )
+        except (OSError, ValueError, UnicodeDecodeError):
+            return []
+
     def _mark_redirect(name: str, form, anchor: str) -> RedirectResponse:
         """Land where the form was: the standalone marks page for a verbatim
         booth (its own index.html cannot show the recorded judgment), else the
@@ -973,13 +1172,29 @@ def create_app(
         place a verbatim-index.html booth can show its marks — that page is served
         untouched by design, so the inline panel never renders there."""
         booth = resolve_booth(name)
-        marks = marks_for(booth)
+        # U4: for a verbatim booth this IS the booth page. `/b/<n>/asks` is a
+        # 308 into here, so the legacy URL records through this call and must
+        # not get one of its own.
+        record_view(booth)
+        held_marks, read_err = hold_read(booth)  # ONE read; see list_booths
+        hold = hold_reason(held_marks, read_err)
+        marks = held_marks if read_err is None else marks_for(booth)
         return templates.TemplateResponse(
             request,
             "marks.html",
             {**base_ctx, "name": name, "name_url": quote(name, safe=""),
              "marks": marks, "marks_open": len(open_marks(marks)),
-             "booth_marks": marks_for_target(marks, None), "marks_page": True},
+             "booth_marks": marks_for_target(marks, None), "marks_page": True,
+             # U4 INV-4, and this page is WHY the invariant needs a third home.
+             # A verbatim booth's own index.html is served untouched, so it has
+             # no Booth-rendered header to carry the lifetime line — this page
+             # is the only surface besides the index card where the Booth owns
+             # the chrome. Without it, the booths most likely to be held (a
+             # report that ASKS something is the archetype) would be the ones
+             # that never say they are.
+             "kept": is_kept(booth),
+             "hold": hold,
+             "expires_in": max(0.0, ttl_seconds - booth_age_seconds(booth))},
         )
 
     @app.get("/b/{name}/asks", include_in_schema=False)
@@ -1002,12 +1217,29 @@ def create_app(
         the whole booth instead of one question at a time.
         """
         booth = resolve_booth(name)
-        marks = marks_for(booth)
+        marks, read_err = hold_read(booth)
-        return JSONResponse({
+        body = {
             "booth": name,
             "marks": [as_dict(m) for m in marks],
             "open": [m.id for m in open_marks(marks)],
-        })
+        }
+        # A DAMAGED file used to come back as an empty list and nothing else,
+        # which is indistinguishable from "you were never asked anything" — and
+        # this endpoint is the ONLY reader a remote session has. Its filesystem
+        # sibling has told the truth since U2: `booth marks` exits 3 on an
+        # unreadable file precisely so a caller can tell "not yet" from
+        # "broken". One question, two surfaces, two answers.
+        #
+        # The STATUS stays 200 and that is deliberate. Reads are lenient here —
+        # the same rule that keeps a poisoned booth from 500ing the index — and
+        # a pinned status code is a promise to remote clients this fix has no
+        # business breaking. The information goes in the body instead: a client
+        # that wants the CLI's exit-3 parity reads `error`, and one that does
+        # not behaves exactly as it does today.
+        if read_err is not None:
+            body["error"] = "this booth's .marks.json cannot be read"
+            body["detail"] = read_err
+        return JSONResponse(body)
 
     @app.get("/b/{name}/view", response_class=HTMLResponse)
     def booth_view_file(request: Request, name: str, f: str):
@@ -1029,6 +1261,14 @@ def create_app(
 
         items = booth_items(booth)
         item = find_item(items, f)
+        # U4: a bookmarked zoom URL is somebody looking — but only once we know
+        # there is an ITEM to look at. Below the 404s, and gated on the record,
+        # because `f` is any path that stats inside the booth: the bug-hunt
+        # panel pointed `?f=.marks.lock` at this and held a booth open with a
+        # file the service created itself. A dotfile is not an item, and a view
+        # of a thing that is not an item is not a view of the booth.
+        if item is not None:
+            record_view(booth)
         marks = marks_for(booth)
         item_marks = marks_for_target(marks, f)
         common = {
@@ -1210,7 +1450,35 @@ def create_app(
     @app.post("/b/{name}/unkeep")
     def booth_unkeep(name: str, next: str = Form("/")):
         # missing_ok: releasing an already-released board is a no-op, not a 500.
-        (resolve_booth(name) / KEEP_MARKER).unlink(missing_ok=True)
+        booth = resolve_booth(name)
+        marker = booth / KEEP_MARKER
+        try:
+            marker.unlink()
+            released = True
+        except FileNotFoundError:
+            released = False              # already released: a no-op, not a 500
+        except OSError:
+            # A `.forever` that is a DIRECTORY raised IsADirectoryError straight
+            # through this route and 500'd it, which made the card's release
+            # button permanently dead for that booth. Pre-existing; the panel
+            # re-exposed it. Removing it is still best-effort, and failing to is
+            # not worth refusing the request over.
+            released = False
+        if not released:
+            return RedirectResponse(url=_safe_next(next), status_code=303)
+        # U4: RELEASE IS ACTIVITY, and now it is a rule rather than an accident.
+        # A released board already survived another full TTL, because unlinking
+        # a file bumps the directory's mtime — behaviour the note above calls
+        # "not intuitive" precisely because nothing declared it. The behaviour
+        # is unchanged; its reason is now stated. Releasing a board is somebody
+        # touching it, so it gets one full TTL, the same as any other look.
+        #
+        # ONLY when something was actually released, which is the correction the
+        # bug-hunt panel forced: an unconditional call made POSTing release at
+        # an already-released booth an endless TTL refresh, contradicting this
+        # route's own no-op promise and diverging from the CLI, which `rm`s the
+        # sentinel without recording anything.
+        record_view(booth)
         return RedirectResponse(url=_safe_next(next), status_code=303)
 
     @app.post("/b/{name}/blur")

booth/marks.py

@@ -217,12 +217,21 @@ 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:
@@ -245,6 +254,8 @@ def _read_raw_strict(booth: Path) -> list[dict]:
     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)
@@ -545,6 +556,34 @@ 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]

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/base.html

@@ -330,6 +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);

booth/templates/booth.html

@@ -1,5 +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
@@ -65,7 +66,7 @@
   {% else %}
     <h1>{{ name }}</h1>
   {% 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 %}{% 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>
+  <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

booth/templates/index.html

@@ -1,5 +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">
@@ -39,7 +40,7 @@
       </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
@@ -69,11 +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>
@@ -111,7 +112,7 @@
       </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
@@ -123,7 +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>
@@ -160,5 +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" %}

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.

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-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-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-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-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-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-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-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

@@ -21,352 +21,82 @@ _As of 2026-09-22:_
 
 - **v1 is gated on seven units** in `ROADMAP.md`, dependency-ordered
   **U1 → U2 → {U3, U4, U5} → U7**, with **U6 independent**.
-- **U1 and U2 are landed and released.** Current version `0.2.2`, deployed to the
+- **U1, U2, U4 and U5 are landed.** U1 `ce598b3`; U2 `c7f9437` → `v0.2.0`,
-  live service, 275 tests green, tree clean, 25/25 booth pages verified 200 after
+  `5e41108` → `v0.2.1`, `026a1fc` → `v0.2.2`; U5 `c015a91` + `95beede` →
-  the deploy. U1 `ce598b3`; U2 `c7f9437` released as `v0.2.0`, then `5e41108` as
+  `v0.3.0`. **U4 landed 2026-09-22** — 396 tests green (341 → 396), deployed and
-  `v0.2.1` (four contract-panel findings), then `v0.2.2` carrying the
+  verified live, 24/24 booth pages 200, layout probe clean.
-  **bug-hunt panel's** nine (below).
+- **U4 released as `v0.4.0`** (operator approved the minor on 2026-09-22).
-- **U5 is IMPLEMENTED and unreleased** as of 2026-09-22. `booth/manifest.py`
+  `c3a97c1` is the unit; the release commit carries the pre-existing fixes the
-  (stdlib-only, INV-1), `.booth.json` per booth, the provenance line on both
+  bug-hunt panel surfaced in touched files. The tag waited for the last gate to
-  index lanes and the booth page header, `--why` / `--title` on `booth new` and
+  close, per the `v0.2.0` lesson — see Tried and abandoned.
-  `booth add`, and the link board + pickup booths announcing themselves as the
+- ⚠ **The 17 consuming handles have NOT been told** that `keep` no longer means
-  service's own. 310 tests, live service restarted, 26/26 booth pages verified
+  "waiting on an answer". That is the one coordination this release genuinely
-  200 and all 26 rendering `unannounced`. **Deliberately NOT tagged yet**: the
+  warrants, and a fleetwide post needs operator approval before it is sent.
-  cold `/heid-contract-review` panel is still in flight and the code-review and
+- **THE NEXT UNIT IS THE OPERATOR'S CALL.** U3 (declared embed seam) and U6
-  bug-hunt gates have not run. That ordering is the 2026-09-21 lesson applied —
+  (benches) are both unblocked; U7 waits on the rest. U6 is independent of
-  a release whose gate is outstanding is premature even when the tier is right.
+  everything and was conceptually unblocked by U5 giving job 5 a home; U3 is
-  Contract: `docs/contracts/u5_booth_manifest.contract.md` (carries its own
+  where verbatim-booth provenance was deferred to, and U4 added a fourth reason
-  seam-review section).
+  to want it — a verbatim booth has no Booth-rendered header, so its lifetime
-- **U5's original framing** (operator, 2026-09-21): **self-announcing booths.**
+  line lives only on the index card and the marks page.
-  `.booth.json` carrying `{handle, title, why, created}`, written by the CLI from
+- **No gate is outstanding.** All three ran on U4 and were folded in: the
-  `$ALTHING_HANDLE`; the index card gains provenance and a one-line purpose, and
+  `/heid-contract-review` panel (`01M34VX0SH23Y3VC92E7GM4S70`), the
-  the index becomes the "what landed" feed the link board was being used as. It
+  `/heid-code-review` panel (`01M34WAFJC3RTERFYBBZJN1SVG`) and the
-  closes job 5 of the five jobs — the one nobody named, and the reason 145 dead
+  `/heid-bug-hunt` (`01M34Y2R0RAJRSN36Q8K4KAB36`). All loops closed with heid.
-  link rows existed. Nothing started: no contract, no blast-radius pass.
+  The U5 round's three are also closed (`01M340PNVRS21HPASZT38PXQPN`,
-- **Two things about U5 are already settled and should not be re-derived.**
+  `01M341E9XAPZEFBSPK9HPGAM0S`, `01M343SXX27Z47C3STXXRC7M42`).
-  (1) `.booth.json` is a DOTFILE, so `booth_items`' existing `startswith(".")` skip
+- **Two dated predictions are pending and must not be forgotten.** U5's adoption
-  already keeps it out of tiles, counts and zips — the same reason `.marks.json`
+  re-measure on **2026-09-29** (two counts, see its entry — already at 3 of 24
-  needed no new exclusion rule. (2) The deterministic-order invariant applies to
+  announced and 2 with a `why`, all from peers told nothing), and the `.forever`
-  whatever U5 adds to the index; the index is ordered newest-first by mtime today
+  re-count **on or after 2026-10-06**, a fortnight after U4 landed, which is
-  and that rule must stay stated. Also worth knowing before scoping: enforcing the
+  U4's success criterion. ⚠ Only 4 booths carry marks at all, so the hold's live
-  link rule without giving job 5 a home first just makes it homeless — that is the
+  blast radius is small and the prediction rests on both halves of U4 — see its
-  lesson from the 69% rot, and U5 is the home.
+  entry for what a null result would and would not mean.
-- **No heid dispatch is outstanding.** The `/heid-bug-hunt` on U2's diff landed
+- **Three methodology proposals from this session sit with the operator**, routed
-  2026-09-22 and shipped as `v0.2.2`; see the dated entry below.
+  by heid rather than decided unilaterally: reshaping the paraphrase gate toward
-- Live service `active` on `:8090`, 25 booths, verified 25 × 3 page types after the
+  a drift-check for narrative-heavy contracts, a standing
-  last deploy. The booth set churns: `sindra20-engines` and `sindra-finalists` were
+  "green-tests-prove-nothing" direction for the code-review gate, and regin's
-  swept during the session, `cr123a-to-d-sleeve` and `sindra` appeared.
+  table-vs-signature consistency pass. They are changes to the `/heid*` skills,
+  not to this repo.
+- The booth set churns hard: 26 → 24 during this session as the sweeper ran.
+  Re-count rather than trusting any number written here.
 
 ## Recent decisions
 
-- `[2026-09-22]` **The U5 bug-hunt panel found a service-wide hang that the
+- `[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`
-  SIZE CAP ITSELF opened — two hours after I added the cap.** `stat` reports
+- `[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`
-  size 0 for a FIFO and 0 for a symlink to `/dev/zero`, so both sail under a
+- `[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`
-  byte cap and then `read_text` blocks with no EOF or allocates until the kernel
+- `[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`
-  intervenes. `list_booths` reads every booth on every `GET /`, so ONE such file
+- `[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`
-  stalls the front page for the whole service with no error and no recovery
+- `[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`
-  short of a restart. Reproduced (`timeout` returned 124), fixed with an
+- `[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`
-  `S_ISREG` check BEFORE the size check in both modules, verified live: the
+- `[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`
-  index answered 200 in 36 ms with two FIFOs planted. **The reusable shape:
+- `[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`
-  `st_size` answers a different question than "can this be read", and a bound
+- `[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`
-  that trusts it inherits everything it does not mean — a hardening fix opened
+- `[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`
-  a worse hole than the one it closed.** Also adopted: the upload path wrote 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`
-  manifest ABOVE its own cleanup guard (4/4), so a failure orphaned a half-booth
+- `[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`
-  whose uniquely-named leaked temp then kept it alive forever; replace-over-
+- `[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`
-  damaged destroyed recoverable bytes (4/4, now QUARANTINED rather than refused
+- `[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`
-  — marks refuse because judgment is not restatable, a booth's description is);
+- `[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`
-  and `booth answer` spelled out its own openness test, disagreeing with
+- `[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`
-  `booth marks` about a partially-answered pick, which is a direct violation of
+- `[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`
-  U2's INV-2. Full triage in `persistent-memory.d/2026-09-22-u5-panels.md`.
+- `[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`
-- `[2026-09-22]` **An existing test stopped me retiring documented behaviour
+- `[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`
-  while fixing a race.** The mtime-restore race is real, and the clean fix —
+- `[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`
-  ignoring a booth directory's own mtime whenever the booth holds anything —
+- `[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`
-  would also have silently retired the rule that RELEASING a kept board resets
+- `[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`
-  its clock, which the CLI header, the README and a deliberately-written test
+- `[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`
-  all pin. That is a TTL doctrine change, not a bug fix. Fixed the concrete half
+- `[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`
-  (a failing `os.utime` used to escape and 500 the route), left the race stated
+- `[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`
-  in the code. **A fix that changes a documented rule is a proposal, not a
+- `[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`
-  patch.**
+- `[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`
-- `[2026-09-22]` **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`.
-- `[2026-09-22]` **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.
-- `[2026-09-22]` **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`.
-- `[2026-09-22]` **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.
-- `[2026-09-22]` **`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.
-- `[2026-09-22]` **`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.
-
-- `[2026-09-21]` **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.**
-- `[2026-09-21]` **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.
-- `[2026-09-21]` **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.**
-- `[2026-09-21]` **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.
-- `[2026-09-21]` **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**.
-- `[2026-09-21]` **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`.
-- `[2026-09-21]` **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.
-- `[2026-09-21]` **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.
-- `[2026-09-21]` **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.
-- `[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]` **Tagging a release while a review gate was still in flight.**
+- `[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`
-  `v0.2.0` was cut and announced to 15 consuming handles; 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`
-  `/heid-contract-review` panel — dispatched BEFORE implementation, as the
+- `[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`
-  discipline says — replied afterwards with three defects in the code that had just
+- `[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`
-  shipped, one of them silent data loss. Nothing about the tier decision was wrong;
+- `[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`
-  the *timing* was. **If a gate is outstanding on the work being released, the tag
+- `[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`
-  waits for it.** The cost was a same-hour `v0.2.1` and a correction note to peers
+- `[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`
-  who had already verified against the broken version.
-- `[2026-09-21]` **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.
-- `[2026-09-21]` **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.
-
-- `[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 = "0.3.0"
+version = "0.4.0"
 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 = [

scripts/booth

@@ -12,6 +12,8 @@
 #   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
 #   booth links                         list the board, numbered, with entry ids
@@ -59,12 +61,28 @@
 # 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
@@ -72,10 +90,12 @@
 # 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

scripts/layout-probe.py

@@ -24,6 +24,14 @@ 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[@]}"`.

tests/test_booth.py

@@ -781,6 +781,10 @@ 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()