fix(cli): check's overlap audit skips scenes the runtime hid - #5758

What

check's content-overlap audit no longer walks inside scenes the runtime has hidden. On a long composition, almost every element sits in a scene that is off screen at any given sample, and the audit was reading styles and geometry for all of them on every sample.

Why

The dense overlap pass seeks to up to 600 times and runs this audit at each one. The audit walked root.querySelectorAll("*"), so its cost per sample grew with the length of the film, not with what was on screen. On a 9-minute composition with about 430 scenes, 7,348 of 7,806 elements were inside runtime-hidden clips at a typical sample, and check ran past its deadline. In a CPU profile of 100 samples, the per-element visibility test in this audit was the largest single cost.

The runtime hides an inactive clip with data-hf-clip-hidden plus inline visibility: hidden. Its rule [data-hf-clip-hidden][style*="visibility: hidden"] * { visibility: hidden !important; } (core timedClipHide.ts) keeps every descendant hidden. So nothing inside such a clip can be on screen, and skipping the subtree changes no finding.

Related work

Part of making check's browser work independent of film length. Bounding the number of dense overlap samples follows in a separate PR.

How

• elementsOnScreenScenes(root) walks with a TreeWalker that rejects runtime-hidden clips, so their subtrees are never visited. It matches the same selector as the runtime rule.
• collectSolidTextBlocks uses it. That is the walk behind __hyperframesOverlapAudit (every dense sample) and the overlap part of the full layout audit.
• An author-hidden ancestor without the runtime marker is still walked, so visibility_escapes_segment behaves as before (its existing test is unchanged and passes).

Test plan

• Unit test never reads inside a scene the runtime hid: the overlap audit returns nothing and never reads styles of an element inside a runtime-hidden clip. With the walk put back to the whole page, it fails (expected [ { …(9) } ] to deeply equal []).
• layout-audit.browser.test.ts passes 3 runs in a row. oxlint and oxfmt are clean.
• Same findings: check on the 9-minute composition, before and after, gives identical layout findings (80 of 80, same code, selector and time). The audit's share of the overlap pass dropped 2-3x. Wall times were taken on a loaded machine and are not quoted.
• Comments follow CONTRIBUTING.md "Comments"