[miguel-heygen]

What

A missing frame file now skips its drawElement self-verify sample on the disk path, as the function's contract says, instead of aborting the render with "ffmpeg or the psnr filter is unavailable". This PR also adds the first tests that run the real disk compare and the real sample arming.

Why

psnrForDiskSample read the file and ran the compare inside one try. A read that failed with ENOENT reached isFfmpegInfrastructureFailure, which reads any ENOENT as a missing ffmpeg binary, so a frame file that was not there failed the render and blamed ffmpeg.

The new tests found that. Before them, nothing ran verifyDiskDrawElementSamples: the tests that mention it mock it out. Making the compare never throw, or arming zero samples in captureDeVerificationFrames, left every test green. Sequential disk capture and parallel workers rely on this check before shipping drawElement frames.

Related work

Refs #7324, whose review exposed the gap. #7255 covers the separate segmented and streaming compare, which already has tests.

How

• parallelCoordinator.ts: reads the file in its own try. A read failure skips the sample; compare failures keep their existing skip or abort classification.
• frameCapture.ts: exports captureDeVerificationFrames so a test can call it.
• parallelCoordinator.verifyDisk.test.ts: real files on disk, psnrDb stubbed, since scoring is ffmpeg's job and the check's logic is what is under test.
• frameCapture-verifyArming.browser.test.ts: real Chrome and a real truth page.

Test plan

• Unit and browser tests added. Fails without the fix: with the read back inside the compare's try, "skips a sample whose file cannot be read, and still checks the next" fails (1 failed, 16 passed).
• Each break is caught, one at a time, then restored: the disk compare never throws (3 failed); arming zero samples (3 failed); comparing against the wrong sample's truth (2 failed).
• New tests: 17 passed, three runs in a row on Linux, with Chrome really running (not skipped). parallelCoordinator.test.ts, frameCapture-verifySampleFractions.test.ts, frameCapture.test.ts: 122 passed.
• Engine typecheck, oxlint and oxfmt on the changed files.
• Manual testing performed: no real drawElement render. It needs a macOS or Windows GPU host.
• Comments follow CONTRIBUTING.md "Comments".

[github-actions]

Edit accuracy: accurate 267 (base branch 267), smooth 114 of those

The gate passes.
Smoothness is reported in the artifact, not gated. A case fails only if it fails 2 of 3 runs.

Quarantined, measured but not gated (8)

• seqplainkeys-none-center-r0-nested-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): fail / pass / pass, passes
• seqplainkeys-none-center-r0-root-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): pass, passes
• seqplainkeys-none-center-r30-root-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): not run
• seqplainkeys-none-pct-r0-root-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): pass, passes
• seqplainkeys-none-pct-r30-root-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): not run
• seqplainkeys-none-px-r0-nested-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): fail / pass / pass, passes
• seqplainkeys-none-px-r30-nested-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): not run
• seqplainkeys-none-px-r30-root-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): not run

[heygengenesis]

Approved at 57e09a5d. No blockers.

• Fix: moving readFile into its own try stops a frame-file ENOENT from reaching isFfmpegInfrastructureFailure, which reads any ENOENT as a missing binary. The skip path now matches the function's documented contract (readFile races are per-sample noise). Compare failures keep their skip/abort classification, and both abort cases are still tested.
• Fails without the fix, checked locally: I restored parallelCoordinator.ts/frameCapture.ts from the merge-base and ran parallelCoordinator.verifyDisk.test.ts. Result: 1 failed / 11 passed, and the failure was the named test "skips a sample whose file cannot be read, and still checks the next". With the PR head: 12/12 passed.
• Tests: the disk tests run against real files with only psnrDb stubbed, which fits since the logic under test is the check, not ffmpeg scoring. The browser arming test is skipped when Chrome is missing. I did not run it locally. CI is green on this head; Windows render verification is still running.