fix(tts): instant pause via cancel+re-speak and resume-from-position via onboundary

[YizukiAme]

Summary

Fixes #249 — Browser TTS pause has ~300ms delay due to speechSynthesis.pause() buffering.
Fixes #250 — Lecture mode TTS resume restarts from sentence beginning instead of continuing from pause point.

Note

This PR depends on PR #253 (fix/discussion-tts-pause) which adds browser TTS pause/resume delegation in useDiscussionTTS. Merge #253 first.

Approach: cancel() + onboundary Word Position Tracking

Both issues stem from Web Speech API limitations. The unified fix:

• Instant pause: Use speechSynthesis.cancel() instead of .pause() — cancel is immediate, no buffering delay.
• Resume from position: Track word boundaries via utterance.onboundary events (charIndex + charLength). On resume, text.slice(lastBoundaryIndex) re-speaks from where the user paused.

Changes

lib/hooks/use-browser-tts.ts (discussion mode, #249)

• speakInternal() — shared by speak() and resume(), attaches onboundary handler
• pause() — sets cancellingForPauseRef = true, calls cancel(), keeps flag true during entire pause period (Chrome fires onEnd asynchronously)
• resume() — slices text from last boundary, re-speaks with same voice via speakInternal()
• cancellingForPauseRef reset moved to speakInternal() to prevent race condition with async onEnd

lib/playback/engine.ts (lecture mode, #250)

• browserTTSBoundaryIndex — tracks charIndex + charLength within current chunk via onboundary
• pause() — slices current chunk from boundary position (not chunk[0])
• browserTTSCurrentLang — saves resolved language to prevent voice switching at language boundaries on resume
• Reset boundary index and lang on new playback, cleanup, and stop

Testing

Scenario	Before	After
Discussion: pause	~300ms audio tail	Instant stop
Discussion: resume	From sentence start	From last word boundary
Lecture: pause	Instant (already used cancel)	Same
Lecture: resume	From chunk start	From last word boundary
Pause at language boundary	N/A	Same voice/language preserved

🤖 AI-assisted and reviewed by Claude Opus 4.6

[wyuc]

Thanks for the PR! I tested this locally and the cancel+re-speak approach does eliminate the ~300ms pause delay, but I ran into a couple of issues:

1. Resume position is imprecise on rapid pause/resume — onboundary only fires at word boundaries, so the resume point can jump back noticeably
2. Punctuation gets spoken aloud — after text.slice(boundaryIndex), the remaining text can start with punctuation, causing TTS to read out "句号", "逗号", etc.
3. Duplicated pattern — the cancel+onboundary logic is implemented independently in both use-browser-tts.ts and engine.ts

The root cause is that Web Speech API doesn't expose precise audio position, so cancel+re-speak is inherently approximate. The original pause()/resume() has the ~300ms tail but gives exact resume position and handles rapid toggling correctly.

These are limitations of the Web Speech API itself, so no easy fix — let's leave this for now and revisit later. Appreciate the contribution!