fix: the preview preloads the coming scene's images, and check stops failing on image swaps

[miguel-heygen]

What

Two image-loading fixes from user reports:

1. The preview loads the next scene's images before the scene starts. Studio's preview serves the images of later scenes as loading="lazy", and the runtime hides a scene that has not started with display: none. Chrome never fetches a lazy image that has no layout box, so the 2 s look-ahead marked the next scene as coming up but its images still did not load: on the scene's first frame the image was missing. The runtime now switches Studio's lazy images in a clip to eager the moment that clip comes within the look-ahead, which is what a paused seek onto a scene already did (both now share one helper). Only the clip's own images switch: a clip nested inside it waits for its own look-ahead, so a chapter's later scenes stay unloaded. Images the author marked loading="lazy" themselves are left alone.
2. hyperframes check stops reporting image loads its own scrubbing cancelled. Check samples the timeline 8 times a second without waiting, by design. In a film that swaps an <img> source through numbered frames, each sample cancels the previous frame's load, and check reported every cancelled request as request_failed ... net::ERR_ABORTED, failing the run. After the sampling, check asks the page about each aborted image: it is reported only when an <img> still shows that URL and its decode fails. A load still pending after 5 s (a deferred lazy image) is not counted as a failure. One a later swap replaced, or that has since loaded, is dropped. A missing file still fails the check, as before (missing_local_asset / http_error). An aborted image that is not an <img> source (a CSS background, an SVG <image>) is no longer reported, since an abort means the page cancelled it, not that it is missing.

Why

• Row 1: a tool or person playing the preview saw the next scene appear without its image until the image loaded, and tools that await img.decode() on such an image waited forever.
• Row 2: check failed every film with an image sequence of non-trivial frame size, though nothing was wrong with the project.

Test plan

• entry.test.ts: a Studio lazy image in a clip within the look-ahead is eager while the clip is still hidden; one further away stays lazy until a seek brings it near; one in a clip nested inside the coming clip stays lazy; an authored lazy image stays lazy. Fails on main (lazy where eager is expected).
• checkBrowser.test.ts: the filter keeps an aborted image only when the page reports it broken, and every non-image failure; a test through the whole check path reports an aborted image still shown and broken, and drops one swapped away and one still shown that decodes; another (fake timers, no real wait) pins that a reload still pending at the 5 s cap is not reported. Fails on main.
• Whole core suite: 4216 passed. CLI typecheck clean; core typecheck clean.
• Playing the preview of a composition whose second scene (in normal flow, starting at 2 s) holds one image, read on the first frame at or after 2 s, 3 runs each:

	main	this branch
image on the scene's first frame	not loaded, loading=lazy (3/3)	loaded, 320 px wide, loading=eager (3/3)

• hyperframes check on a 4 s film swapping one <img> through 60 PNG frames at 30 fps:

	main	this branch
2.7 MB frames	exit 1, 27 × request_failed ... net::ERR_ABORTED	exit 0, none
4 KB frames	exit 1, 1 × request_failed ... net::ERR_ABORTED	exit 0, none
an <img> pointing at a file that does not exist	reported	still reported (missing_local_asset), exit 1

Before / After

The preview of a 4 s composition whose second scene (in normal flow) starts at 2 s and holds one 1280×720 image, played from 0 and captured on the first frame at or after 2 s (read at t = 2.00–2.01 s, screenshot done by t ≈ 2.2–2.3 s). The image is served 1.5 s late to stand in for a real network. 3 runs per side, same result each time.

Before

main (156e87766c7): scene 2 has started and its image has not loaded (complete: false, loading=lazy).

After

This branch (905473da374): the same frame shows the image (complete: true, 1280 px wide, loading=eager), fetched during the look-ahead.

Not in this PR

• A tool that awaits decode() on every image of the preview will still wait on images of scenes far beyond the look-ahead: keeping those unloaded is the preview's memory saving. Tools that need every image should load the capture document (?hf-capture=1), which serves no lazy images.
• The deprecated validate command keeps its own request listener unchanged. It seeks only 5 times with a 150 ms settle each, not a dense no-settle pass, so it is much less likely to cancel its own loads.

[github-actions]

Edit accuracy: accurate 2059 (base branch 2059), smooth 1463 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 at exact head 905473da3749f2b2d9a28d8fae64ef93b9c2badc. packages/core/src/runtime/init.ts:388-396,2782-2786,3868-3889 scopes eager promotion to Studio-marked images in the approaching clip and shares it with paused seeks; authored lazy images and nested later clips retain their timing. packages/cli/src/utils/checkBrowser.ts:374-411,414-441 checks aborted image requests after scrubbing while retaining HTTP errors and other network failures.

A Chromium reproduction using actual base/head bundled runtimes, a production-positioned head script, and a synthetic 320px PNG delayed by 1.5 seconds observed the later clip’s first visible frame with image incomplete (width 0) before and complete (width 320) after, in 3/3 runs per revision. This confirms the mechanism, not the PR body’s exact Studio fixture or screenshots; those hosted screenshots were not independently inspected. Focused runtime tests passed 32/32; all 11 required checks passed at this head. The focused CLI suite could not collect locally because of unbuilt transitive workspace packages.

Verdict: APPROVE
Reasoning: The first-frame preload and own-scrub cancellation paths have no verified correctness regression in the reviewed diff.

— Review by tai (pr-review)