fix(cli): fold canvas pixel hashes into the frozen-sweep fingerprint

[vanceingalls]

What

Folds a canvas-pixel hash into the frozen-sweep audit fingerprint (__hyperframesLayoutGeometry in layout-audit.browser.js). Each visible <canvas> is downsampled to 8x8 and its pixels hashed into the per-sample signature; tainted/zero-sized/unreadable canvases hash to a constant.

Why

Wild report: strict check reported sweep_static for a seek-correct, pixel-only canvas animation (native 4K WebGL PNG-sequence render) — a transparent moving DOM sentinel was needed to make it pass. The frozen-sweep guard's fingerprint was geometry+opacity of DOM elements only, blind to a canvas repainting without any element moving.

Validation

Built a minimal pure-canvas-motion fixture (zero DOM motion, only a canvas repaint) and confirmed: WITHOUT the fix, check reports sweep_static (bug reproduced); WITH the fix, it doesn't. Also confirmed no new false positive on a DOM-motion fixture. CLI build clean.

🤖 Generated with Claude Code

[vanceingalls]

• fix(cli): surface video frame-injection failures in snapshot output #2255
• fix(shader-transitions): pin cloned scene roots to composition dimensions #2254
• fix(cli): fold canvas pixel hashes into the frozen-sweep fingerprint #2253 👈 (View in Graphite)
• chore(producer): resync stale committed puppeteer-core symlinks to lockfile version #2259
• fix(producer): disk-headroom gate before hdr raw frame pre-extraction #2256
• feat(cli): preview --browser-no-gpu for gpu-unstable hosts #2252
• fix(cli): settle post-seek font subset loads before audit screenshots #2251
• fix(engine): densify drawelement self-verify with parallel worker count #2250
• fix(producer): withhold de parallel router below a machine ram floor #2249
• fix(producer,engine): gate drawelement on ancestor background-image + tail verify sample #2247
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[miga-heygen]

#2253 — fold canvas pixel hashes into the frozen-sweep fingerprint

Verdict: LGTM 🟢

Fixes false-positive sweep_static for WebGL compositions where motion is pixel-only (no DOM elements move). canvasPixelHash downsamples each visible canvas to 8x8 and folds a simple hash (hash * 31 + data[i]) into the geometry fingerprint as c:${hash}.

Key safety properties: tainted/zero-sized/unreadable canvases hash to "x" (constant), so they can't introduce false NEGATIVES for DOM-motion comps — they're no worse than the current code where canvas content is invisible to the fingerprint. The try/catch around getImageData handles cross-origin tainted canvases gracefully.

Review by Miga

[miguel-heygen]

Reviewed current head 17db454. Canvas hashing is bounded to an 8x8 readback, visibility-filtered, and fail-silent for tainted/unreadable canvases without weakening DOM-motion detection. Required checks are green. APPROVE.

[james-russo-rames-d-jusso]

Reviewed at 17db454.

Right diagnosis — a canvas repainting without any element moving is exactly the class the geometry+opacity fingerprint was blind to, and folding a downsampled pixel hash into the same per-sample signature is the least-invasive way to close it. try/catch around drawImage covers tainted/zero-sized/unreadable canvases with a constant hash — no worse than pre-PR for those cases, no new false negative for DOM-motion comps.

One thing worth thinking about before landing:

Question — WebGL preserveDrawingBuffer: false and readback stability. ctx.drawImage(canvas, 0, 0, 8, 8) on a WebGL canvas reads the current composited pixels. Chrome guarantees this for WebGL contexts created with preserveDrawingBuffer: true (or read within the same event-loop turn as the GL draw call). For preserveDrawingBuffer: false (the default), the browser MAY clear the drawing buffer between composites — subsequent drawImage reads can return blank or the last-composited-frame depending on state.

For a WebGL comp that renders every frame (your positive-repro case), each sample is fresh — hashes differ across seek samples where the app draws different content, hashes match where it doesn't. Works.

For a WebGL comp that draws ONCE and stops (splash/still-frame): the front-buffer state at sample N vs N+1 could differ purely because of buffer preservation, not because pixels changed. Risk case: same static canvas across two samples produces hash A then hash "x" (or A then constant-black), and the fingerprint diff false-positives motion → misses the sweep_static detection it was meant to catch.

Vance validated the positive (pure-canvas motion detected) and one negative (DOM-motion still detected). The specific "static WebGL canvas, single draw, subsequent seeks" negative isn't in the validation notes. Worth spot-checking that one before rollout, since misclassifying static-WebGL as motion would flip the sign of the fix for that authoring pattern. off.getContext("2d").drawImage(webglCanvas, ...) is likely stable across the same event-loop turn as the seek → screenshot cycle, but the seek→sample gap may cross a composite boundary.

LGTM from my side.

— Review by Rames D Jusso

[jrusso1020]

Approving per the channel stamp-flow: Rames-D and Magi both cleared, CI clean at 17db454.

The base branch was changed.