Address correctness-review findings; migrate component tests to render()

[trevormunoz]

Addresses three findings from a correctness review (verified against source before accepting), plus a reviewer-supplied end-to-end regression test.

Changes

fix(sync) (60d50d9) — production bug. videoController ran the near-end backoff (duration - 1) after the lower clamp, so any duration shorter than the 1s backoff produced a negative seek (duration 0.5 → -0.5; duration 0 → -1). Now re-clamps the lower bound afterward. New tests cover short/zero durations and a [0, duration] invariant sweep; the prior tests only used 60s durations.

test(sync) (b1bc631) — guard tests. Three syncMachine guard tests named behavior they never exercised (one asserted state topology instead of sending a scroll; one had no assertion; the annotation-unchanged case never sent a second update). Rewritten to await the mediaDriven → ready return and then send the competing event. Calibrated: breaking either guard now fails the matching test.

test (699c249) — unmount-leak fix, done idiomatically. The suite mounted components with raw svelte.mount() (260 calls, 1 unmount), so Svelte destruction never ran between tests — effects, listeners, and transcript polling leaked across the suite. Switched to render() from the already-installed vitest-browser-svelte, whose beforeEach(cleanup) auto-unmounts (the idiomatic Svelte 5 browser-mode tool, not a hand-rolled registry). Net −744 lines as mount/target/teardown boilerplate and the createChildSnippet/createContextCapture helpers go away.

• Adds TestContextHarness / TestRootContextCapture / TestTranscriptContextHarness and a small TestTwoTranscripts wrapper (context goes through a wrapper because Svelte's createContext key can't be injected via render()'s context map).
• Adds Root.unmount-lifecycle test: a VTT resolution completing after unmount must not write transcript state. Calibrated by awaiting the exact loadVTTTranscript promise, after an earlier microtask-drain version produced a false green.
• Integrates element-scroll-seek test: scrolling the real custom element's transcript to the bottom seeks the audio to the last cue end (29s), not media duration (40s). Closes a gap where disabling the scroll listener passed every prior test.
• Root.transcript-generation: replaces a 100ms wall-clock wait with vi.waitFor (flaky under full-suite load; behavior unchanged).
• Two Root tests with a genuine render()-vs-mount() async-fetch race are left on mount() + explicit unmount() (documented; no leak).

Verification

• pnpm test: 745 passed across 70 files
• pnpm typecheck: 0 errors
• pnpm lint: eslint + prettier clean
• Each finding fixed test-first and calibrated (breaking the guarded code fails the test)

No public API or demo changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_015KfYGrLi7iJoevyVQsn5ur