The layout audit fingerprints each element by rect and opacity, plus an
8x8 pixel hash for canvas and video. A composition whose only change over
the timeline is font-variation-settings touches none of those, so all six
samples hash identically and the audit reports sweep_static on a timeline
that is visibly animating.
Duplexed variable fonts make it unavoidable rather than unlikely. Recursive
holds one advance width across its whole weight axis by design, so even the
element rect stays fixed while the letterforms change.
Fold the computed font-variation-settings into the per-element fingerprint.
It reads normal on every element that does not use the property, so no
existing composition's fingerprint changes.
This is the third instance of this failure class in the same function,
after the canvas-repaint and opacity-reveal cases whose regression tests
already sit alongside these. Two more join them: a scene where only the
axes move must change the fingerprint, and a scene where nothing moves
must hash identically twice. The second is the load-bearing one, since a
fingerprint that varies on its own would make sweep_static unfireable and
every green layout verdict meaningless.
A search miss already reached PostHog, where nobody was watching it. It now
also goes to the same channel a CLI rating goes to, by the same best-effort
POST, so a move the catalog is missing is read rather than queried.
The forward is bounded and swallowed, and the ack prints before it, so a gap
report can never fail or delay the command that sent it.
* fix(cli): persist what the server answered for a captured page
A page that renders is not a page that succeeded. An error page has a title,
a palette, typefaces and a DOM, so every extractor downstream reads it happily
and produces a design system belonging to whoever wrote the error page rather
than to the site's owner.
The status was already read, once, to feed `detectBlockedPage`, and then
dropped. That helper cannot stand in for it: it decides whether the rendered
document LOOKS like a protection wall, over a minimal-DOM heuristic that a
rich error page passes. "Was this blocked?" and "what did the server answer?"
are two questions, and widening the first to carry the second would leave a
heuristic owning a fact.
So the response status is persisted plainly, as its own record, and every
consumer decides for itself what a non-success response means for its product.
Written before the blocked-page check runs, so the record's absence means
"navigation never produced a response" — a third state distinct from a status
of 404 and from a status of null.
* fix(cli): report the response status in the capture command's JSON output
The status reached `CaptureResult` and was then dropped at the CLI boundary,
which is the same read-once-and-discard that made the error page harvestable
in the first place. `--json` is the documented programmatic surface, and an
agent reading `ok: true` off a capture of a 404 has no way to see it there.
Bare `hyperframes skills` passed `--skill '*'` to the upstream installer, which discovers every SKILL.md in the clone, so a full install put 26 skills on disk against the 20 the manifest publishes (the six repo-internal ones under .claude/skills / .agents/skills). It now resolves the canonical manifest and installs its names, the same path `skills update` uses; offline it warns and degrades to the pinned core set instead of the wildcard. Docs point the full set at `npx hyperframes skills` and say what `skills add --all` really pulls.
Rebuilding footage means checking the render against the source at the same timestamps. `--against <video>` saves the reference frame at every captured time (frame-exact seek: -ss after -i) plus a labelled render|reference pair sheet built with the existing contact-sheet helper. Reference frames and pair sheets are derived artifacts, kept out of the snapshot list like contact-sheet.jpg. No new dependencies.
* fix(fonts): make Google Fonts subsetting CSS text-transform aware
Extends the subset character closure to cover locale/context-sensitive
case transforms and non-case CSS text-transform values:
- Parse lang attributes from authored HTML and apply toLocaleUpperCase/
toLocaleLowerCase for each detected locale (covers Turkish İ/ı,
Azeri, German ẞ, and other locale-dependent casing)
- Map ASCII U+0021–U+007E to fullwidth equivalents U+FF01–U+FF5E when
full-width appears in the source
- Map small hiragana/katakana to full-size equivalents when
full-size-kana appears in the source
- Preserve the existing 1700-char encoded URL budget and full-font
fallback
Closes#3496
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* refactor: extract helpers to reduce complexity and duplication
Split extractGoogleFontsText into addCaseClosure, addFullwidthVariants,
and addFullSizeKanaVariants. Extract subsetTextFor test helper to
eliminate repeated URL→text boilerplate.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: validate lang tags, scope transform gates, deduplicate closure input
- Validate lang attributes with Intl.getCanonicalLocales before passing
to toLocaleUpperCase — malformed tags (en_US, x, 123) no longer throw
RangeError.
- Gate fullwidth/kana expansion on text-transform declarations instead
of raw html.includes — a CSS class named .full-width no longer eats
half the URL budget.
- Deduplicate characters before the closure loop (new Set) to avoid
redundant locale conversions on base64-heavy compositions.
- Benchmark now asserts the transform cost delta, not just that one
small fixture fits under the cap.
- Restore over-approximation comment.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: prevent regex bridging across rules and add case-insensitive matching
Exclude {} from the text-transform regex character class so the match
cannot cross rule boundaries when the trailing semicolon is omitted.
Add /i flag so uppercase declarations (text-transform: FULL-WIDTH)
are not missed. Test now uses a fixture with both a text-transform
declaration and a .full-width class to exercise the bridging case.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: remove redundant \s* to eliminate quadratic backtracking
Drop the \s* between : and [^;{}]* — the character class already
matches whitespace, and the adjacent quantifiers caused quadratic
backtracking on inputs like "text-transform:" + " ".repeat(N).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: replace regex with linear indexOf/slice scan (js/polynomial-redos)
The text-transform regex backtracks O(n²) on input with many
text-transform: runs and no ;{} between them. Replace with a linear
indexOf/slice scan following the repo's existing pattern for this
CodeQL rule (beatFile.ts, utils.ts, parseStoryboard.ts).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Agents were shipping <video src="….jpg"> / <img src="….mov"> because the existing URL-resolvable check stayed green. Report an error so the agent rewrites the tag; skip audio, <source>, and extensionless URLs.
* fix(core): catch PostCSS parse errors instead of dropping compositions
Invalid CSS in a sub-composition style block made postcss.parse throw
inside scopeCssToComposition. The throw propagated to the composition
loader's catch block, which emptied the host — silently dropping the
entire scene. Lint swallowed the same error via catch { continue },
reporting 0 warnings.
Two fixes:
- Runtime: wrap postcss.parse in try/catch and return the original
(unscoped) CSS on failure, so the composition still mounts
- Lint: emit a css_parse_error finding instead of silently continuing
Fixes#3585.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: drop unparseable CSS instead of leaking it unscoped
Return "" on PostCSS parse failure so sub-composition stylesheets
that cannot be scoped are dropped rather than injected unscoped into
the parent document. Updates test fixture to use valid+malformed CSS
that demonstrates the leak risk.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
declaredSize() matched the rel token "apple-touch-icon" exactly, so the
legacy "apple-touch-icon-precomposed" spelling (one token, not two) fell
through to 0 instead of the 180px Apple default, even though the selector
already collects it. Same page, one spelling apart, opposite winner inside
tier 1 — it never drops a candidate and never beats the .ico tier.
Switch to a startsWith check on the token. Exact-token matching stays for
mask-icon, where a longer rel really would be a different asset.
* fix(producer): degrade gracefully when font cache directory is unwritable
When the font cache root (~/.cache/hyperframes/fonts/) cannot be created
(EPERM on read-only filesystems, restricted home directories, etc.), the
render aborts with a raw mkdir error. The cache is an optimization, not
a requirement — a missing cache should mean slower first renders, not
broken renders.
Fall back to a temporary directory under os.tmpdir() when the configured
cache root fails, so Google Fonts downloads still proceed. The fallback
cache is per-process and not persistent across renders, but the render
completes.
Fixes#3412.
* fix: use mkdtempSync for font cache fallback, restore warning
Rames Jusso's review caught a regression in the force-push: the
predictable tmpdir path is unsafe (symlink attacks in world-writable
dirs), and the warning log was dropped. Restore the mkdtempSync
pattern matching lambdaFontCacheRoot, add a CLI hint per Miguel's
request, and reuse the ephemeral root across calls.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: emit font cache fallback warning once per run, not per font
Gate the warning on whether this is the first fallback activation.
The ??= already suppresses repeat mkdtempSync, but the warn fired
for every font family in the composition.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(cli): capture the best declared favicon, not the first one
Pages routinely declare a legacy 16px .ico first and the good asset (an SVG,
or a 180x180 apple-touch PNG) after it. The capture's page evaluate kept only
{rel, href} and the download loop took whichever candidate fetched first, so
the .ico won on every such page. The dropped sizes/type attributes are the only
evidence of quality: page.html on disk does not keep the <link> tags and only
the winner's bytes are fetched, so the choice was unrecoverable downstream.
Keep sizes and type, and rank candidates before downloading: SVG first, then
the largest declared size (an unsized apple-touch-icon counts as 180), then
.ico. The loop still falls through to the next candidate when one fails to
download, so the ranking changes which icon wins, never whether one lands.
Ranking is a pure function over the declared attributes, unit-tested against
the link shapes three sites actually publish.
* fix(cli): rank a pinned-tab mask-icon below real favicons
`link[rel*="icon"]` also matches `rel="mask-icon"`, Safari's pinned-tab
asset: a single-colour silhouette drawn in a browser-chosen tint, not the
site mark. It is served as an SVG, so ranking by format alone promoted the
outline above the page's actual colour favicon.
Give mask-icon its own lowest tier rather than dropping it, so a page that
declares nothing else still lands an icon instead of none.
Capture drops assets for four reasons and reported none of them, so a folder
with thirty images and a folder truncated to thirty images were the same
object. Every drop site was a bare `continue`, `break`, `return null` or an
empty `catch`, and the only signals downstream were two hand-written warning
strings that fired when the budget was already gone before a download pass
started, which is the one case where the pass could not say how much it lost.
`downloadAssets` and `downloadAndRewriteFonts` now return an `AssetDropCounts`
tally beside their result, incremented at the single line that performs each
drop. `capture --json` carries it as `dropped`; the human summary prints a
`Dropped:` line when it is non-zero.
Four reasons, three decisions and one failure:
size-floor fetched, then judged too small to be a real asset
budget-exhausted the post-navigation clock ran out before this one
cap-reached 30 inline SVGs, 30 fonts, or 6 faces per family
unavailable the request or the write failed
A break now counts everything it did not reach rather than the one it stopped
on, because "how many did we lose" is the question and one is never the answer.
The two budget warnings are gone. Both existed only to cover the case where
the budget ran out before a pass was called, so both passes are now called
unconditionally: a zero budget makes each loop break on its first item and
record `budget-exhausted` for the rest, which costs no network and produces a
real number instead of the word "some". The single remaining warning is derived
from the tally, so the prose and the count cannot disagree.
Measured on a live capture of a large marketing site:
default budget 232 kept, 91 dropped (39 size-floor, 20 cap-reached,
32 unavailable)
15s budget 30 kept, 299 dropped (279 budget-exhausted, 20 cap-reached)
Same page, same command, and until now both runs described themselves the same
way.
* fix(producer): transport safe extraction failure metadata
* refactor(producer): generalize public error metadata
* test(producer): use vendor-neutral media hosts
The ENAMETOOLONG regression writes 150 real clip files and the mixer
existence-checks each one: ~58ms on Linux, but past vitest's 5s default on
the Windows lane. packages/engine sets no global testTimeout, so heavy
tests here carry an explicit one.
On timeout its abandoned async work kept calling the shared runFfmpegMock
after afterEach cleared it, so the next test saw 5 calls instead of 3 and
lost its queued once-implementations to the leak. mockReset stops an
aborted test from handing leftovers to the next one.
* feat(registry): say what a promoted template's slots mean, not where they sit
Every declaration described its slot by position — appName was "small credit in the corner
of every banner" — and every editing contract described the supplied brand only by what it
may not touch, chatgpt-exchange going as far as "only the subject discussed inside the
conversation". An agent filling those slots has nothing telling it that the video is an
advertisement for that brand, and remixes shipped with an invented brand instead of the
real one.
Adds `portrays` to the 35 identity-bearing variables across seven templates: what the slot
stands for, as a list, kept orthogonal to `role` (which says which aspect of the
composition a knob affects, not what a value means to a viewer). Rewrites those variables'
descriptions to state meaning, and gives every TEMPLATE.md the affirmative sentence it was
missing. notes-reveal now records honestly that it declares no identity slot at all, which
is why a remix of it cannot carry a brand.
`portrays` is optional metadata: rendering ignores it, the promoted contract still projects
{id, type, role}, and packages that predate it keep working. Variable ids, types, roles,
labels, and defaults are untouched, so installs and renders see the same contract.
* feat(notes-reveal): give the template somewhere to say whose ad it is
This was the one promoted template that declared no identity slot at all, and the
previous commit could only record that honestly. A remix of it could not name the
advertiser, could not carry its domain, and had nowhere to place its mark — so the
ad it produced was unattributable no matter how good the copy was.
The closing card ends at y=1708, leaving 212px of dotted scene. A sign-off strip
lives there: `brandLogo` (portrays subject_logo) above `brandDomain` (portrays
subject_domain), in the card's own plum and Courier Prime. Putting it outside the
card means the hand-lettered headline and the checklist keep the geometry they were
designed with — verified by `hyperframes check`, which reports the same 0 errors and
the same warning counts as origin/main, with one more contrast check (367/367 AA).
No name slot: this template has no text slot that can hold a bare brand name.
`cardTop`'s marker line continues into the fixed words "OF ONE FILE", and a name set
beside a wordmark reads twice — rendering it proved that before this was committed.
The mark plus the domain carry the attribution, which is what the other end-card
templates do.
Rendered to check rather than assumed: the packaged default reads as the HyperFrames
wordmark over hyperframes.heygen.com, and a resolved symbol mark (Notion's cube over
notion.com) sits correctly in the same strip.
* fix(notes-reveal): make the note the ad, and stop the headline leaving the card
Three defects, all visible in the reported gamma remix.
The seven body lines carried their characters as pre-split `<span class="ch">`
elements so the reveal animation could address them by id. That made the note
permanently unwritable: 46% of the 24.9s runtime, and the only place the note
says anything, was fixed copy about making videos with HyperFrames -- shipped
inside an advertisement for a Dutch DIY chain. The spans are now built at
runtime from seven declared variables, keeping the same ids, so the animation
is untouched and the copy is the brand's.
`cardMid` and `cardBottom` ("OF", "ONE FILE") were literals with no variable at
all, so a remix could only ever change the first of the card's three lines.
Every text slot sat at a fixed left edge on one line, guarded only by a
character cap -- the wrong unit. `cardTop` allowed 12 characters; "THE POWER"
(9) fits and "CREATE DECKS" (12) does not, which is why both reported gamma
remixes shipped with the closing headline over the edge of the card. Slots are
now fitted to their box at render time, shrinking type only as far as needed and
never past the designed size. The card headlines also carried
`data-layout-allow-overflow`, switching off the one check that would have caught
it; those waivers are gone. The note's title and body keep theirs, because
`#note-body` is translated across the whole scene and leaves the canvas by
design.
Timing is preserved rather than regenerated: each body line is revealed inside
the frame window the packaged rhythm already used, so `JUMPS` and the hand-tuned
`SCROLLS` curves stay valid and the composition is still 24.867s for any length
of copy. A line whose length matches the packaged copy keeps its rhythm frame
for frame.
Verified: rendered with the packaged defaults, SSIM 1.000000 across all 746
frames against the previous template -- the design is bit-identical. Rendered
with the gamma values that overflowed, and with a stress set of long and
one-word values: both stay inside the card and the note, both still 24.867s.
`hyperframes check` passes, with layout warnings down from 28 to 8 and no new
issue class. Note that `check` takes no variable values, so it guards the
template's defaults and the fit is what guards a remix.
* fix(notes-reveal): make the fit floor and the character caps agree
A cap that admits a string still too wide at the floor font size puts the text
back outside its box -- the exact bug the fit was added for. With the floor at
0.6 and the title cap at 40, a 40-character title still overflowed `#note-body`
by 30px, because "Things nobody told me" is 21 characters and 900px at 88px, so
42.9px per character, and 40 characters need to reach about 0.45 of the design
size to fit.
Floor to 0.45 and the caps in: titles 34, body lines 52, cardTop and cardBottom
16. Every cap is now satisfiable with room to spare, and the common case stays
far from the floor.
Found by probing all eight promoted templates with values about 1.8x the length
of their packaged defaults, which is the length a real remix produces; short
sentinel values only ever prove a slot is wired.
Defaults are unaffected: SSIM 1.000000 across all 746 frames, still 24.867s,
`hyperframes check` still passes.
* fix(exchange): scroll the composer to follow the caret
`prompt` is an editable slot in both templates, but `.comp-text` is a
fixed-width single line with `white-space: pre` inside a composer that clips,
and the packaged prompt already reaches the mic -- it has no headroom at all.
A longer prompt therefore slid under the mic and the send button and lost its
tail, with the caret hidden behind them: the remix read as having typed into
nowhere, cut mid-word.
Both now scroll the typed text left to keep the caret in view, the way a real
single-line input does, on the very per-character times the reveal already uses.
The limit is measured to the opaque round button rather than the mic, because the
packaged prompt's caret already sits 2px past the mic's left edge in
chatgpt-exchange -- measuring to the mic would scroll the packaged composition.
Shrink-to-fit was the other option and it is worse: with no headroom it shrinks
the packaged prompt too, and it still cannot absorb a prompt twice as long
without unreadable type. It was tried, measured, and dropped.
A prompt that fits emits no keyframes and the element is never handed to GSAP at
all -- even a zero translate stamps a transform, which changes how the text
rasterizes.
claude-exchange puts this inside its existing `build()`, which already re-runs on
`document.fonts.ready`, so it measures the final font. chatgpt-exchange builds
once, so its scroll keyframes are kept in a list and rebuilt on the same event.
Verified on the packaged defaults: claude-exchange SSIM 1.000000 across 1284
frames. chatgpt-exchange is not frame-deterministic -- rendering it twice
unchanged gives 0.999959 with 249 frames under 0.9999, and before-vs-after gives
0.999967 with the same 249 -- so the change sits inside its own noise. Durations
unchanged at 14.9s and 21.4s. `hyperframes check` reports byte-identical results
to the unchanged templates.
* fix(notes-reveal): fit the checklist cells instead of counting their characters
The three checklist rows were the only text slots with no width fit, so
their character cap was doing the job a measurement should: it was set to 8
because that was the widest the packaged strings needed, and a gamma.com
remix delivered "DIRECT MEE" and rendered "DIRECT M".
Fit the labels and values like every other text slot -- dividing the
label's scaleX(0.86) back out of its column, since fitWidth measures the
untransformed width -- and return the cap to a backstop at 18.
"DIRECT MEE" needs no shrink at all: it ends at x=694 against a tick at
x=715, so the truncation was never about room. Measured: the packaged
defaults still render at their authored 44px and 50px, and all three values
clear the tick.
* fix(chatgpt-exchange): nest the caret scroll so its rebuild can be undone
`tl.to()` returns the TIMELINE, not the tween it created, so `caretScrollTweens`
held 49 references to the master timeline. The fonts-ready rebuild then called
`.kill()` on the master timeline 49 times and re-added its keyframes on top of
the ones it meant to replace. Measured with an overflowing prompt:
master timeline detached from the global timeline (tl.parent -> null)
scroll tweens on #cge-typed: 49 before the rebuild, 98 after
49 timestamps left with two competing tweens on the same property
GSAP's `Animation.kill()` interrupts and detaches rather than killing children,
so the composition kept rendering under an explicit seek -- which is why the
packaged-defaults SSIM check passed and never saw this. It is still wrong: the
timeline is off the ticker, and the stale keyframes were measured against the
fallback font.
Only an overflowing prompt reaches it, i.e. exactly the remix the scroll exists
for. The packaged prompt emits no steps, so nothing is created and nothing is
killed.
Now one nested child timeline holds the steps: a real object that can be killed
and replaced, and killing it cannot reach `tl`. Children sit at the same absolute
times and the nest is added at 0, so the motion is unchanged.
Verified: overflowing prompt holds at 49 scroll tweens across the rebuild with no
duplicated timestamps and the timeline stays attached. Packaged defaults against
the pre-feature baseline -- SSIM 0.999998 over 894 frames, 13 frames below
0.9999 against a 249-frame noise floor for this template, duration 14.9s
unchanged.
Found by Magi in review of #3562.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* feat(lint): catch a timeline's return value being treated as a tween
Magi asked for a regression that makes the caret-scroll failure unable to return.
A render fixture cannot be that guard, and neither can a unit test. Both were
measured before choosing this:
- The unit-test DOM has no layout. In happy-dom `getBoundingClientRect()` is
{left: 0, width: 0} and offsetWidth/offsetLeft are 0, so the template's own
guard computes limit = -8 and returns before collecting a single keyframe,
whatever the prompt length. `document.fonts` does not exist there either, so
the rebuild never runs. Both preconditions are unreachable -- the test would
pass on the broken code and the fixed code alike. Rames flagged this; verified.
- A render fixture would not fail either. Rendering the long prompt through the
real pipeline on the broken vs fixed template gives SSIM 0.999994 over 894
frames with ZERO frames below 0.999. The renderer drives the composition by
explicit seek, and the duplicated keyframes are added last so they win, which
is exactly why `.kill()` on the master timeline stayed invisible. A golden mp4
would have cost a fixture and a shard and guarded nothing.
What is testable is the root cause, and it is a static property: `tl.to()` returns
the TIMELINE, `gsap.to()` returns a Tween, and the two read identically. So the
guard is a lint rule, which needs no DOM, fires on the source, and covers every
composition in the registry rather than this one template.
The rule keys on the timeline handles `collectTimelineVarNames` already finds,
which is what keeps `gsap.to()` and `Array.from()` out of it. Pushing a timeline
return into an array is flagged outright -- that is the collect-to-discard shape.
A bound return is flagged only once something tween-scoped is aimed at it, so a
pointless-but-harmless capture stays quiet.
Verified against the real files, not just synthetic ones: the pre-fix
chatgpt-exchange raises exactly one error and names the offending line; the fixed
file is clean. Across all 642 HTML files under registry/ and examples/ the rule
finds nothing, and lint output on the eight promoted blocks is identical to
origin/main error-for-error and warning-for-warning, so error severity costs no
existing template. 540 lint tests pass, 6 of them new.
Requested by Magi in review of #3562; unit-test DOM measurement by Rames Jusso.
* fix(capture): let vision captioning authenticate the way a server can
Three defects in one phase, all of which end with a capture that reports
"Captioned N/N images" and then "0 images captioned with Gemini" — a
successful-looking run that hands the agent nothing to see by.
1. Credential. The captioner only accepted an API key. A server deployment
holds a service account, not a key, and a rejected key is indistinguishable
from an unset one here: every request returns empty text and no error. Vertex
is now a first-class provider, ranked above the bare key and below an explicit
OPENROUTER_API_KEY opt-in, configured by HYPERFRAMES_VERTEX_SERVICE_ACCOUNT +
HYPERFRAMES_VERTEX_PROJECT_ID (region via HYPERFRAMES_VERTEX_LOCATION). It
carries its own model default because the Gemini API's flash-lite preview id
is not resolvable on Vertex.
2. Empty captions. Thinking tokens are drawn from maxOutputTokens, so a model
left free to think can spend the whole budget and return no text — a
successful request with no caption. thinkingBudget is pinned to 0; a one-line
factual caption needs no reasoning.
3. Native abort. Rasterizing a batch of SVGs concurrently drove up to SVG_BATCH
simultaneous librsvg renders through libvips and corrupted the heap:
`free(): unaligned chunk detected in tcache 2` (SIGABRT) during this phase,
twice in fourteen days, losing the whole capture each time. A native abort
cannot be caught, so the concurrency is removed rather than handled —
rasterization is serialized and libvips' worker pool is bounded, while the
vision requests, which are the slow leg, stay parallel. Throughput barely
moves: 225 captions across three real captures, 0 failures, 13-25s each.
* test(capture): pin the rasterization loop to one render at a time
The serialization fix shipped without a regression test on the grounds that native
heap corruption is not unit-testable. The corruption is not, but the property that
prevents it is: `sharp` is mocked to record how many renders are in flight, and a
six-SVG batch must never reach two. A deliberately slow caption stub makes
overlapping renders the faster path, so a future refactor that "optimises" the loop
back to `Promise.all` fails here instead of aborting in production.
Also covered: `sharp.concurrency(1)` is applied — serializing the loop while leaving
libvips' pool at the host core count still fans one render across every core — and an
unrasterizable SVG is skipped without breaking serialization for its siblings.
Verified as a real guard: reverting only contentExtractor.ts to origin/main fails 7 of
the 22 cases in this file.
* fix(capture): tell the truth in the asset-descriptions header when Vertex captioned
The provider gate in `contentExtractor` accepts Vertex when a project and a
service account are both set -- which is the configuration a server
deployment actually has. The header written next to the captions still
tested only for an API key, so a capture whose captions Vertex had just
generated was labelled "GEMINI_API_KEY not set -- descriptions below are
catalog-derived".
That header is not cosmetic: it travels into the context the template
editor reads, telling it to distrust captions that are real.
Mirror the same two variables here, and name every provider in the fallback
text instead of only the API key.
* fix(capture): hand libvips' worker pool back after the renders
`sharp.concurrency(1)` is process-global and was set once, for the whole life of
the process. The bound is right for the rasterize loop -- a native abort in
libvips cannot be caught, so the renders must not overlap -- but its scope was
every later sharp caller in the process, none of which asked for captioning, all
of them pinned to one thread from then on.
Now the host's value is read first and restored in a `finally` around the
rasterize loop, so a skipped SVG cannot cost the process its threads either. The
vision requests below are network work and gain nothing from a pinned pool.
The mock had to grow the getter half of sharp's API -- `concurrency()` with no
argument reports the current value -- since save-and-restore is untestable
without it. Verified as a real guard: dropping only the restore fails both new
cases.
Raised by Rames Jusso in review of #3561 and concurred by Magi.
A card centered at y=.860 can still cover the painted V2A pill. Intersect the element's getBoundingClientRect with the keepout instead of testing whether its center sits inside the band.
heygen-rui-bot