Root causes

Two consecutive main Windows runs timed out in the first checkBrowser.test.ts case (29227777037, 29220733386). The test mocked @hyperframes/core/compiler, but production reaches bundling through bundleWithLocalizedFonts, so the first case cold-loaded the unrelated producer/font graph under the full Windows monorepo workload and exhausted Vitest's 20s ceiling.

After the Studio/Ular cutover merged, both Linux and Windows then consistently failed two useTimelinePlayer assertions. #2291 intentionally narrowed downgrade preservation to enriched sub-composition children so undoing a split can remove a top-level clone instead of leaving a ghost clip; the old tests still expected every missing top-level element to be preserved.

Fix

• Mock the browser-check unit's actual bundleWithLocalizedFonts boundary and assert the project directory crosses that seam. The real helper remains covered by its own integration test.
• Align the timeline merge regressions with the new policy: sub-composition children (including anonymous same-label clips) remain preserved, while a dedicated regression proves a missing top-level split clone is dropped.

No production timeout or behavior was loosened.

Verification

• bun run build ✅
• CLI browser/font suites: 12 passed ✅
• Studio timeline-player suite: 28 passed ✅
• CLI + Studio typecheck ✅
• Changed-file format + lint: 0 warnings/errors ✅
• 10 repeated isolated checkBrowser.test.ts runs: 10/10 passed, 1.79–2.35s wall time ✅
• First PR Windows rerun: CLI 1,649 tests passed with the original timeout gone; the run then isolated only the two stale Studio assertions corrected by the second commit. Render-on-Windows passed end to end. A fresh full CI/Windows run is attached to the updated head.