* feat(registry): add 25 image carousel blocks (5 families × 5 variants)
Five carousel families, each with 5 style variants:
- Orbit (1–5): image cards on a spinning 3D Fibonacci sphere
- Path (1–5): cards following animated CSS motion paths
- Circle (1–5): circular carousel layouts
- Vision (1–5): Apple Vision-style spatial presentations
- Text Circle (1–5): circular carousels with text overlays
All blocks are 1920×1080 at 6s, with 12–24 configurable image slots.
Includes catalog preview thumbnails for each block.
Co-Authored-By: Jake Moran <jake.moran@heygen.com>
* refactor(registry): host carousel block images on the CDN
The registry is served straight out of this repository
(DEFAULT_REGISTRY_URL points at raw.githubusercontent.com), so every byte a
block ships is permanent history. The 25 carousel blocks added 421 JPEGs,
27.7 MiB in a checkout, and made the diff 472 files. Only 23 of those images
were distinct: the same 12-24 placeholders were copied into every block.
files[] entries gain an optional `url`. When set, the installer fetches the
bytes from there instead of joining the registry base. `path` does not change
and still says where the file lands relative to the item, so composition HTML,
target mirroring and `hyperframes add` behave exactly as before.
Keys are content-addressed, so the 396 manifest entries resolve to 23 objects,
and a changed image gets a new URL rather than a stale one cached behind
`immutable, max-age=31536000`.
The catalog preview renderer copies an item's directory and renders it, so it
needs the same materialisation step. Without it the preview draws every card
blank and reports success, which is worse than failing.
Also drops registry/catalog/, 25 hand-made thumbnails referenced by nothing;
catalog previews are rendered by CI and served from docs/images/catalog.
Verified: all 23 objects return 200 from the CDN with hashes matching their
keys; `hyperframes add carousel-orbit-1` against a local registry installs 24
real JPEGs; the preview render produces the album art, and produces blank
cards when the fetch step is removed.
* style(registry): format the carousel composition HTML
`oxfmt --check .` covers the whole tree, and these 25 files were never run
through it. The pre-commit hook only formats staged files, so nothing local
caught it.
* feat(catalog): publish the carousel blocks without republishing their images
These 25 blocks had no Catalog page. Every other item in the registry has one,
so they shipped invisible: installable by name, unfindable by browsing.
Generating them naively undid the change they were added by. The Catalog
payload copies an item's assets into docs/public/, which is tracked, so the 396
images this PR just removed came back as 43 MB one directory over — worse than
the 3 MB they started as, because each block got its own copy.
The copy exists because these compositions assemble `img.src` at run time out
of a variable value, so there is no `src="..."` in the markup for the payload's
asset scan to resolve. An unpredictable path can only be satisfied by serving
every file beside it, which is what `needsOwnDirectory` asks for.
An absolute URL needs no directory: the scan already skips any `https:`
reference. So for the payload path only, hosted files are left undownloaded and
the composition's variable defaults are rewritten to their URLs. The preview
renderer still downloads them, because it paints real frames and a missing file
is a blank card.
The explorer posts every value to the preview frame on mount, including
untouched ones, so the page's variable list carries the URLs too. Left as local
paths they would have overridden the payload's own defaults and asked the frame
for a file that was deliberately never published.
Result: 25 pages, 25 payloads, zero bytes of image added.
Verified: a spike item declaring no assets at all rendered its 24 covers from
the CDN, proving the variable-default path; payload generation for a carousel
block now writes no item directory and no shared asset; the preview render
still produces the album art. mint validate and mint broken-links pass on the
new pages. test:scripts is green.
* refactor(catalog): split the hosted-asset step out of prepareProjectDir
Two functions rather than one: finding the composition and rewriting its
variable defaults are separate jobs, and inlining the mode branch pushed
prepareProjectDir past the complexity gate it was already sitting on.
Behaviour is unchanged. Re-verified both paths after the split: the payload for
a carousel block still writes no item directory and no shared asset, and the
preview render still produces the album art.
* feat(catalog): give the carousels their own shelf
25 image carousels landed in Showcases and were 53% of it, so the scenes that
shelf exists for disappeared underneath them. That is the same shape the 24
editor themes made, and it gets the same fix they got.
Keyed on the first tag, which is this file's stated grouping rule, rather than
on the name. `screen-flow-carousel` leads with `product-demo` and stays on the
shelf that says what it is for; a future carousel that is not named
`carousel-*` still lands here.
Showcases 47 -> 22, Carousels 25, and no existing item changed shelf.
* fix(registry): centre the circle-5 carousel path in its composition
Its ring was centred at x=3832.6 in a 3840-wide composition, so it sat on the
right edge and most of it fell outside the frame. Only a few cards were ever
visible, cropped, with two thirds of the composition empty.
The exported path carried absolute coordinates from a layout that was never
recentred. Shifting the four vertices by (-1912.597, -4.340) puts the ring on
the composition centre. Handles are relative, so only the anchor points move
and the shape is unchanged.
carousel-text-circle-5 shares the identical path and had the identical fault.
The other three circle variants sit within 12% of centre, which reads as
authored placement rather than the same bug, so they are left alone.
* fix(catalog): rebuild the circle-5 payloads after recentring the path
The Catalog preview plays the payload, not the composition on disk, so
recentring the source changed nothing a reader sees. The payload still carried
the old vertices and the ring still hung off the right edge of the frame.
Verified the consumer this time, not just the producer: both payloads now
resolve to a path centre of x=1920. The other 23 rebuild byte-identical, so the
formatting pass did not reach them.
* feat(catalog): promote Carousels to its own section
It was a shelf inside Scenes & demos, which is where a scene type belongs by
kind but not by weight. At 25 items it is larger than Data & charts (17) and
Blocks (13), each of which is already a section holding a single shelf, so the
catalog's own precedent puts it one level up.
Pulling it out also takes the largest section in the catalog from 120 items to
95, which is the reason the shelf was added in the first place.
The two circle-5 pages change because their embedded source block carries the
recentred path; nothing else in them moved.
---------
Co-authored-by: Jake Moran <jake.moran@heygen.com>
Co-authored-by: Miguel Angel Simon Sierra <miguel.sierra@heygen.com>
* test(producer): pin unpinned drawElement stall retry
* fix(engine): bound drawElement frames so a wedged renderer falls back instead of failing (#3172)
* fix(engine): bound drawElement frames so a wedged renderer falls back instead of failing
A single drawElement frame could kill an entire render. On one comp
(caption-editorial-emphasis) drawElementImage returns normally and the renderer
then stops draining its task queue: the setTimeout(…, 0) that schedules
toDataURL never fires, the capture page.evaluate never settles, and 60s later
the stage watchdog fails the whole render. Deterministic on 152.0.7977.30,
always the same frame. Root-cause detail in PRINFRA-488.
Two gaps, both closed here.
1. Nothing below the 60s stage watchdog bounded a frame. Adds a per-frame
deadline (HF_DE_FRAME_TIMEOUT_MS, default 15000, 0 disables) around the WHOLE
frame operation, not just the drawElementImage call — the stall surfaces at
whichever page round-trip comes next, and was observed at both the capture
and the seek's background-image decode. Bounding one call missed it.
Deliberately NO per-frame screenshot fallback for this error: once the
renderer stops scheduling it is wedged for every subsequent round-trip on
that page. Measured — the screenshot fallback blew the same deadline. The
frame fails fast instead, so the recovery that actually works can run.
2. That recovery was ineligible. shouldRetryViaPinnedFallback only retried a
generic capture failure when the worker count was PINNED by the inversion or
the router; a comp that engaged drawElement on the ordinary single-worker
path had no whole-render fallback at all. A renderer stall is now retryable
on any routing — the failure is a property of drawElement itself, and the
retry re-renders on a fresh page via screenshot. Reported as
de_fallback_reason "de_renderer_stall" so it is distinguishable from
capture_error in telemetry, and counted per-session as deFrameTimeouts.
Verified end to end on the repro: previously exit 1 with "stalled: no frame
progress for 60000ms"; now the deadline fires, the render retries via
screenshot, and RENDER_OK with a valid 1920x1080 / 240-frame / 8.0s MP4.
Tests cover the predicate both ways plus the cross-package error match. Engine
1,481 and producer 587 green (audioPadTrim.integration flakes only under
full-suite parallel load — passes in isolation, and on clean HEAD it is green
too, so it is not from this change).
Refs PRINFRA-488
* fix(engine,producer,cli): count drawElement frame timeouts where they happen
`session.deFrameTimeouts` was incremented inside `captureFrameCore`'s catch,
but the deadline races that function from OUTSIDE it — `withFrameDeadline` is
applied by `captureFrameToBuffer`, and `DeFrameTimeoutError` is constructed
nowhere else. So the branch was unreachable: on a wedged renderer the counter
stayed 0, the engine-side stall log never printed, and
`CapturePerfSummary.deFrameTimeouts` reported 0 on every affected render. The
end-to-end recovery worked throughout — `isDeRendererStallError` matches by
name at the producer — only the observability was dead.
Moved the increment and the log into an `onTimeout` hook that fires from the
deadline itself, deleted the unreachable branch and the now-unused
`isDeFrameTimeoutError`, and threaded the count out to
`render_completed.de_frame_timeouts` so the rate is graphable apart from the
other `capture_error` fallbacks. A stall always costs a whole-render re-run,
which is why it deserves its own series.
Also corrected the `fallbackReason` docstring, which still listed only
`psnr | blank | oom | capture_error` after this branch added
`de_renderer_stall`.
`withFrameDeadline` is exported for a two-case fake-timer test: the deadline
fires and calls `onTimeout` once, and a resolving inner promise passes its
value through with no late timer. The first case fails if the counter moves
back inside the work promise.
(cherry picked from commit ec2b2b5b2a)
* fix(render): retry sequential screenshot stalls
* fix(render): bound parallel worker capture phases
---------
Co-authored-by: Vance Ingalls <vance@heygen.com>
* feat(lint): catch leftover marker heads and guessed marked shafts
Dash-draw on a marked path shows the arrowhead before the shaft exists.
A long marked path that misses every node in both user and screen space
is the same detach as the CTM-paste bug, without the counterfactual.
* feat(check): flag orphan connectors and unbalanced style tags
Catch a visible shaft while fewer than two nodes are on stage (enter-early /
exit-late), and extra </style> that dumps CSS onto the frame.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(check): judge connector_orphan by the shaft's own endpoints
The rule counted anchors stage-wide and skipped whenever two were on. That
tier is every visible text-bearing or opaque element under 15% of the
stage, so a title and a footer alone satisfied it — the check could not
fire on a composition that had any chrome, which is all of them.
Endpoints are now resolved per shaft, against candidates gathered by
layout instead of by visibility, so a node hidden with opacity:0 still
owns the endpoint that meets it and its visibility is what the rule tests.
An endpoint that meets no node is left to connector_detached.
* revert(lint): drop the marker rules this repo would only duplicate
marker_orient_typo and marker_dash_draw_on already run in Zephyr's Python
connector lint, down to the same regexes and the same message and fixHint
strings. Zephyr concatenates its local findings onto this bridge's result
with no dedupe, so shipping them here reports each one twice and counts it
twice, and every later edit has to land in two places.
The browser-check work stays: a marked shaft that meets no node, and an
orphan shaft, are both runtime facts this repo can see and a source-reading
rule cannot.
* chore(check): drop the comments added with the orphan rewrite
* fix(check): score orphan and detach on overflow-opted connector layers
Allow-overflow is a clip opt-out, not a lifecycle skip. Orphan now uses
the same connector names as detach so flow and arrow shafts are judged.
* fix(check): see axis-aligned shafts, and script tags closed with a space
connector_orphan gated on isVisibleElement, which rejects anything whose
bounding box has no height or width. A straight horizontal or vertical
connector is exactly that, and it is the common shape, so the rule was
blind to most of what it exists to catch: across 47 production
compositions it produced nothing. It now tests paint -- display,
visibility and the opacity chain -- and leaves size to the 80px chord
floor it already had. The same 47 compositions now yield three findings,
each naming the endpoint that is still dark.
unbalanced_style_tags stripped scripts with a regex that only matched
`</script>`. `</script >` is valid, so the script survived the strip and
a `"</style>"` string literal inside it counted toward the tag balance,
reporting an error on a composition whose tags are paired. Both the strip
and the closer count now tolerate whitespace before the `>`, which is also
what CodeQL flagged on this branch.
* fix(lint): count style tags in one pass instead of stripping scripts
CodeQL reads the `source.replace(/<script.../)` as an incomplete HTML
sanitizer. It was never one — the rule only needs to not count `<style`
tokens that live inside a script string. A single alternation scan, where
the script branch consumes the whole block, gets the same counts without
a replace.
* test(check): cover the two orphan guards that no fixture was crossing
Neutering `shaftIsPainted`'s opacity check, or the endpoint proximity
threshold, left the suite green. Both are false-positive guards: a
connector staged hidden before its reveal, and a hidden element that
happens to be the nearest thing to an endpoint without belonging to it.
The hidden-shaft case now runs over all four ways a shaft can be
invisible, so the display and visibility clauses are crossed too.
* test: pin the two boundaries raised in review
A closer inside an html comment is already dropped upstream by
stripHtmlComments, so the rule never sees it. The test fails if that
stripping regresses; adding a comment branch to the rule's own scan
would have reintroduced the polynomial-redos pattern that stripper
exists to avoid.
A node hidden with display:none has no geometry, so nothing can place it
near an endpoint and connector_orphan stays silent. connector_detached
does not pick it up either, since it allows a half-attached shaft by
design. That case is uncovered, and the fixture says so.
* fix(check): type the hidden-shaft fixture table as a style record
* fix(check): key connector_orphan by geometry and let a live node win the endpoint
Two defects found in review.
connector_orphan was persistence-tiered but absent from the geometry key,
so several id-less shafts orphaning at one sample each collapsed into a
single finding that then read as held rather than transient. On the 47
corpus compositions this recovers 9 findings across 5 compositions where
3 across 2 were reported; one composition fades its nodes and its
connectors on a shared stagger, so four shafts each outlive a different
node and only one of the four survived the collapse.
The endpoint scan took the nearest candidate and asked whether that one
was hidden, without asking whether a visible box was also in range. A
staged halo sitting on the node it belongs to produced a finding naming
the halo. A visible candidate within threshold now settles the endpoint.
Fixtures added for both, and for the two guards a mutation sweep found
unpinned: the connector-shape test and the dash-offset skip.
* test(check): kill every surviving mutant in the connector_orphan path
Seven guards still passed the suite when deleted: the four candidate
filters, the connector layer's own visibility gate, the defs/marker skip
and the chord floor. Each now has a fixture where that guard alone
decides the outcome.
The earlier sweep that missed them was mutating the wrong copy of a
shared line — connectorAnchorRects and connectorEndpointCandidates carry
the same filter text, and a first-match replace edited the detached one.
Mutations are scoped by enclosing function now.
* test(check): order the halo before the node it hides under
The halo fixture proved nothing. Both boxes contain the endpoint, so both
score gap 0, and the strict tie-break keeps whichever comes first in
document order — which was the visible node. The old code picked the same
winner and stayed silent too.
Putting the halo first makes the tie-break hand it the slot, so only the
visible-candidate check keeps the endpoint attached.
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Combines two fixes to the layout audit:
- `staticIssueKey` no longer includes `issue.text` for `content_overlap`
and `text_occluded`; both builders now emit a unique selector for both
ends, so the selector pair is the identity. An animating subject
(count-up, typewriter, rotating word) no longer splits one held
collision into transient groups.
- A clipping box that is its own nearest constraint reports
`clipped_text` only, instead of also reporting `text_box_overflow`.
A non-clipping self-constraint still reports.
Combines two fixes to the motion sidecar path:
- An ambiguous selector no longer aborts the whole spec. It is reported
as a finding, the assertions that depend on it are skipped and named,
and every other assertion still evaluates. `assertionTargets` is now
exhaustive over the assertion kinds so a new kind cannot silently fall
back to an arbitrary first match.
- A sidecar that will not parse is reported as a spec finding instead of
ending the run, so the composition is still audited.
* fix(cli): clarify publish visibility and claim links
* style(cli): apply oxfmt to publish visibility test
* chore(skills): regenerate skills manifest for updated references
* fix(cli): stop an in-place re-publish claiming it made the project private
* test(cli): pin in-place visibility copy to the plain re-publish route
A brand band shows logo artwork on a transparency checker, where a badge (a
mark knocked out of a full-bleed disc or square) reads as a solid blob. The
capture kept only the first icon that downloaded, so which of a site's icons
survived was decided by ranking alone, and nothing recorded what the survivor
actually looked like.
Keep them all. Each declared icon is written as assets/icon-<rel>-<sizes>.<ext>
and classified as bare-mark, badge, or unknown, then the best one is copied onto
the historical assets/favicon.<ext> stem so a stem match still finds it.
Classification is deterministic and cheap. SVG is read from markup, not
rendered: these files routinely fill via CSS custom properties behind a
prefers-color-scheme query, and librsvg resolves neither, rasterising one real
favicon to a fully transparent image while Chrome draws it correctly. Raster
icons are sampled at the midpoint of each edge rather than the corners, because
a full-bleed disc has four transparent corners and would otherwise read as a
bare mark.
The headline prefers a bare mark, then falls back to the existing ranking.
Preference is binary on purpose: an "unknown" is not promoted above a badge,
or an undecodable .ico would outrank a good SVG purely for being unexaminable.
extracted/icons-manifest.json records every icon, its shape and the evidence
for that shape, plus which file became the headline, its rank, and why. A
choice whose losers are invisible is indistinguishable from having had no
choice, which is how a substituted icon went unnoticed in the first place.
Worth noting for consumers: sites often declare no bare mark at all. An
apple-touch-icon is composited onto an opaque tile per Apple's spec, so it is
a badge, not a mark. The manifest says so rather than implying the preference
was satisfied.
A capture is one session with two halves: Chrome navigates the page with a
browser User-Agent, then Node fetches the assets that page referenced. Those
halves sent three different identities — "HyperFrames/1.0" from the asset and
media downloaders, a bare "Mozilla/5.0" from the stylesheet inliner, and the
real Chrome UA from the navigation itself.
An origin is free to answer those differently, and anti-bot edges do. Capturing
one large site, GET /favicon.svg answers 403 text/html to "HyperFrames/1.0" and
200 image/svg+xml to the UA the very same capture had just navigated with. The
favicon ranker had already picked that SVG as the best declared icon; the 403
discarded it and the downloader fell through to the next candidate, so the icon
written to assets/ was chosen by the CDN's bot rules rather than by the ranker.
The capture reported it as one "unavailable" drop and carried on.
Hoist the navigation UA into CAPTURE_USER_AGENT and use it for every
out-of-band fetch the capture makes: favicons, images, og:image, fonts,
stylesheets, Lottie JSON and videos. One constant is what stops the two halves
drifting apart again.
Verified end to end against that site: before, assets/favicon.png (the
apple-touch icon) plus one unavailable drop; after, assets/favicon.svg, byte
identical to the file the site itself serves.
* fix(studio): prevent preview hang on burst external file rewrites
Two interacting bugs caused Studio to freeze when multiple processes
(generator, check, snapshot) burst-wrote index.html within seconds:
1. SSE listener leak: the /api/events handler added a watcher listener
per client connection but never removed it on disconnect. Reconnects
accumulated dead listeners, each triggering readFileSync on every
file change and writing to closed streams.
2. Generation starvation: processChange incremented generationRef and
awaited drainPendingChanges. A second event arriving mid-drain bumped
the generation, causing the first drain to bail at the generation
check. With rapid writes, no drain ever completed and Studio stayed
frozen on stale content.
Fix 1: use stream.onAbort() to remove the watcher listener when the
SSE connection closes.
Fix 2: gate processChange with a draining ref. While a drain is in
progress, stash the latest event. On completion, process the stashed
event — the last write in a burst always completes its reload.
Closes#3646
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(studio): align coordinator tests with drain serialization
Update existing test to expect the new behavior: when two events
fire in quick succession, the first drain completes and triggers a
reload (previously it was silently discarded). The stashed event
then starts a second drain.
Also fix the burst-write test to use the drains array pattern and
explicit act() flushes for stashed event processing.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(studio): stash events with allowDuplicate and harden listener cleanup
Address Rames's review findings:
- Stash with allowDuplicate: true so re-dispatched events are not
swallowed by the duplicate guard (the identity was already written
on the way in, so the stashed event matched itself on re-entry).
- Wrap SSE keepalive loop in try/finally so the listener is removed on
both abort and throw, not just abort.
- Restore stale-completion guard test coverage lost in the rename.
- Use await act(async () => {...}) for burst dispatches so assertions
depend on the stash guard rather than scheduling.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(studio): simplify drain serialization and restore SSE cleanup
Restructure processChange into intake + drain loop:
- processChange is now synchronous — validates, dedupes, checks own
echoes, enqueues the accepted payload, and starts the drain loop
- startDrainLoop runs while the pending slot is non-null, draining
one event per iteration via drainOnePending
- No recursive void processChange(...) from finally, so no
allowDuplicate escape hatch needed — stashed events never re-enter
intake guards
SSE listener: restore stream.onAbort alongside try/finally. Hono's
sleep() never throws, so finally alone doesn't fire on disconnect.
Both paths call removeListener (Set.delete is idempotent).
Tests: remove stale-drain test that contaminated subsequent tests by
emptying the shared roots array mid-test. Use sync act() for burst
dispatches — the stash decision is synchronous.
All 10 coordinator tests pass locally (NODE_ENV=test).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
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.
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(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.
heygen-rui-bot