fix(cli): check finishes on video-heavy projects, its geometry seeks skip only video-frame waits - #5715

Reviewed HEAD 4bad251c0feb610c29a1e5731a94f6974347c732 against effective base d05bc4215d5757da86956217f4a8a4b2ab11b1fd.

[P2] Keep asynchronous DOM-source completion in dense geometry seeks

packages/cli/src/capture/captureCompositionFrame.ts:42,344–351

The opt-out skips the whole shared seek barrier, not only video decoding. The public registerFrameSource contract accepts any Element and an asynchronous render callback. The real adapter registers source.seek(sourceTime) with this same barrier (frameSources.ts:130–143); its queue serializes asynchronous draws and coalesces later requests. renderSeek dispatches adapters synchronously but does not await those DOM mutations. The dense pass now reads text geometry before that work finishes.

Concrete source-derived scenario: a 2s, 30fps, data-no-timeline composition has two opaque, absolutely positioned text blocks. A deterministic registered source waits 500ms, then puts them together only for source times 0.45–1.2s. With sparse --at 0,1.875, both ordinary audit samples are separate. BASE waits for each dense source draw and observes overlap at 0.5, 0.625, 0.75, 0.875, 1 and 1.125s: six observations spanning 625ms, which the existing persistence rule promotes to an error. On HEAD, a legal 150ms dense traversal finishes before the first asynchronous draw; reads see the previous non-overlapping pose, intermediate seeks coalesce, and all six findings disappear. Later completion cannot repair samples already collected. Restoring only the completion wait restores the observations; a synchronous DOM renderer also remains correct with the opt-out.

Please retain synchronization for asynchronous DOM-producing frame sources while bypassing unnecessary media-frame completion. The new fake-page test proves that the barrier call is skipped and the default still waits, but it never executes the runtime/source queue or reads geometry, so it does not protect this contract.

Validation and scope: source-only review, including both changed files in full (BASE before HEAD), the actual source queue/barrier/seek dispatch, dense driver and overlap reader, and persistence/report paths; an independent adversarial pass corroborated this finding. The schedule above is a manually authored finite model, not a browser reproduction or measured benchmark. No repository tests, browser, native tools or CI were run. Unrelated adapters/platform execution remain outside this semantic audit; the author's performance measurements were not independently reproduced.

Verdict: REQUEST CHANGES
Reasoning: Avoiding video decoding during text-geometry sampling is reasonable, but bypassing all completion work also introduces a reachable false-negative path for supported asynchronous DOM sources. Preserve that geometry contract without reinstating unnecessary video-decode waits.

— tai

Reviewed HEAD 404856287b4c409ff2e5406bc8879561706463b6 against effective base 41486b9725ab2aae4f457d75dcd625931d103bb3.

The previous asynchronous DOM-source finding is repaired: dense seeks now invoke the shared completion barrier, excluding only classified native-video/proxy-frame holds. Frame-source readiness/draws, synchronous hf-seek registrations, VFX capture and image-held seeks remain generic completions; ordinary audit and inspected export capture retain full waits. Skipping video holds does not remove them from a later full barrier.

Audited: all eight changed files in full on both effective BASE and HEAD (BASE first), the actual fixing patch and public review history, plus bounded production caller/producer/consumer integration. Traced CLI option serialization to the installed runtime hook, dense geometry sampling to its text-rectangle reader, frame-source queue/abort ownership, GPU heartbeat interaction, retained failures/concurrent drains/reset, native/proxy cleanup and default engine capture. Both independent bounded source-audit lanes completed; no concrete introduced reachable blocker survived.

Trusting / limits: untouched repository behavior outside that bounded integration audit, actual browser/native decoder behavior and platform performance are not independently validated. Read the regression tests, including the real frame-source adapter/barrier case and the serialized CLI callback case, but did not execute tests, CI, browsers or benchmarks. The author's timing/platform results are not independently reproduced.

Verdict: APPROVE
Reasoning: The selective video-frame opt-out preserves the asynchronous DOM/GPU completion contract that the earlier whole-barrier bypass broke, while default capture behavior remains wait-all in the inspected production paths. No blocking introduced defect found in this source-only review.

— tai