[miguel-heygen]

What changes for a user

Nothing directly: this adds a check. The 0.8.148 fix (#5458) for a talking-head video that went black while Studio preview played is now guarded by a real project played in Studio's real preview, not only by unit tests. If the preview ever again forces a script-hidden backdrop visible, CI goes red.

Before (the #5458 visibility change reverted)

After (main)

What the check does

packages/studio/tests/e2e/fixtures/aroll-backdrop is a 10 s, 9:16 film: a looping A-roll video (a copy of the repo's own colour-bar design-panel-qa/assets/test.mp4, 28 KB), an untimed full-frame #plate above it that the film's own GSAP script shows only from 6 s to 8 s, a caption on top, and two sub-compositions with GSAP timelines. In Studio preview the runtime stamps #plate with full-length timing so the timeline can list it, which is the shape that broke.

aroll-backdrop.mjs plays it in Studio and reads the preview's pixels and the element on top at the frame's centre: the video must show paused at the start, while playing at about 1 s, and paused after the backdrop's window; the backdrop must show while playing inside 6-8 s. Every wait is on a condition (Play enabled, the video has a frame, the playhead inside a range). Studio can re-fit the preview just after a pause, so a screenshot taken across that reflow is retaken (up to 5 times).

It runs in the existing studio-timeline-viewport job, in the same step as the trim-snap check (ci.yml), which also runs on packages/core/** changes.

Checked

• With only fix(core): an element its script hides stays hidden while studio preview plays #5458's visibility skip in packages/core/src/runtime/init.ts reverted (runtime rebuilt): exit 1, playing: A-roll hidden (chroma 4, on top "plate", plate visible).
• On main: exit 0, three runs in a row, no failures.
• oxfmt, oxlint, comment checks clean.

Not covered here

• Run on Linux headless Chrome for Testing only; not yet on the CI runner's Chrome (this PR's CI run is that check).
• It does not fail if a future Studio stops stamping the backdrop at all; it records that in its output.

[github-actions]

Edit accuracy: accurate 2061 (base branch 2061), smooth 1640 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 (0)

[terencecho]

Reviewed 46e930b3c10db8aae09024cd1adffb59361f8e20 against effective merge-base 943b6512b43b106c79839a977f3bfa4aa115a11d (the observed API base tip was e1925d4317fcd13c0e5799cd4767a3e2dd4c9238). Full scope: seven files, +369/−1; one modified workflow, five added text files and one added MP4. No supported introduced blocker found after the first-pass and independent adversarial source audits.

Audited

• All changed text, the fixture's complete asset references, and the real Studio preview/player/compiler/runtime caller chain. The MP4 is byte-identical to the existing tracked design-panel-qa/assets/test.mp4 (28,273 bytes; SHA-256 4662cef1ee4423640d4db8b8880ea889d6e0af6e4466d88f5ee15f2dc6d18030). That establishes provenance, not independently decoded pixels.
• The untimed root-child #plate reaches Studio runtime stamping. The scripted visibility window is [6,8); the runtime seeks the timeline before applying timed visibility during playback. An offline source-extracted control removing only the runtime-stamped visibility skip produces the intended outside-window failure; it does not revert unrelated #5458 behavior.
• packages/studio/tests/e2e/aroll-backdrop.mjs:119–154 checks actual iframe screenshots and composition hit-testing: A-roll before the plate during playback, a positive dark/topmost plate witness inside its window, and A-roll in paused samples before and after. A plate that never appears cannot satisfy the positive witness.
• .github/workflows/ci.yml:178–183,849–866,885–886,1005–1037 routes this change to the existing Studio job, builds the relevant packages, copies the fixture, runs this check against the server built from the job's checked-out PR tree, and propagates its failure through the accumulated exit status. The normal PR checkout can be the integration merge tree; this is source wiring, not a claim of a bare-head CI run.

Non-blocking coverage limits
The walk pauses before the post-window screenshot and does not exercise an explicit seek. It records rather than asserts the runtime-stamped marker, and the fifth screenshot is accepted even if the rounded frame bounds still differ. A still-playing post-window sample, seek sample, and stability-exhaustion failure would strengthen coverage; none establishes an introduced failure in this test-only delta. CI currently keeps this check's JSON in logs without configuring its optional PNG output.

Verification / trusting
The offline, source-extracted guard/oracle controls passed with simulated DOM, clocks, media and pixel values. I did not run Puppeteer, decode media, build Studio, query/rerun CI, or verify the author's repeated native-pass/fail claims. No submitted GitHub reviews or inline findings were present in the refreshed history; the edit-accuracy bot comment is not a review approval. Approval is on the audited code merits, not those execution claims.

— tai

Verdict: APPROVE
Reasoning: The test reaches the production preview path and has both negative and positive visual oracles for the targeted backdrop regression, with source-confirmed CI failure propagation. The remaining gaps are coverage hardening, not demonstrated blockers.