feat(voice): tts playback via Yapper (phase 2)

[dimakis]

Summary

• TTS module (lib/tts.ts): text chunking at sentence boundaries, synthesis fetch with AbortSignal cancellation, singleton AudioContext management, WAV playback via AudioBufferSourceNode
• useVoice extension: ttsAvailable from health poll, ttsEnabled/selectedVoice with localStorage persistence, lazy voice list fetch on first enable, speak() with sequential chunk synthesis, stopSpeaking() with abort
• VoiceSettings component: speaker toggle (hidden when TTS unavailable), voice picker dropdown grouped by language
• ChatView wiring: auto-speak on new assistant message (tracked by messageId ref, not array length), stop playback on user send
• CSS: speaker toggle with active/speaking pulse, compact voice picker

Addresses all review feedback from #106: singleton AudioContext, AbortSignal on synthesize, messageId tracking instead of messages.length, sequential playback (no pipelining for MVP), lazy voice fetch, dynamic default voice.

Test plan

• 34 new tests across 4 files (tts, useVoice TTS, VoiceSettings, ChatInput mock update)
• Full suite passes: 579/579
• Lint: 0 errors
• Manual: enable TTS toggle → send message → hear response spoken
• Manual: speaker toggle hidden when Yapper offline / TTS model not loaded
• Manual: stop playback when sending new message
• Manual: change voice → next response uses new voice
• Manual: long response → chunked synthesis

🤖 Generated with Claude Code

[dimakis]

PR Review: feat(voice): tts playback via Yapper (phase 2)

Clean implementation overall — good singleton AudioContext, proper AbortSignal cancellation, solid test coverage (34 tests). A few things to address:

🐛 Code blocks read aloud (ChatView.tsx)

The auto-speak effect joins all blockType === 'text' blocks, but these can contain markdown code fences. TTS will read raw code aloud — import React from react semicolon.... Should strip fenced code blocks (and possibly inline code) before passing to speak().

⚠️ No length guard on TTS (ChatView.tsx)

Very long assistant responses (multi-paragraph explanations, large diffs in text blocks) will produce huge synthesis requests. Consider a max character limit with a truncation notice, or at least chunking at the ChatView level before delegating to speak(). The chunking in tts.ts handles synthesis-level splitting but doesn't guard against sending a 10k-char wall to begin with.

📝 Minor: MIN_FRAGMENT_LEN not in constants (tts.ts)

TTS_CHUNK_MAX_CHARS is in constants.ts but MIN_FRAGMENT_LEN (10) is hardcoded in tts.ts. Inconsistent — either move it to constants or document why it's different.

🧪 Missing test: ChatView auto-speak effect

The auto-speak useEffect has non-trivial logic (messageId tracking, running guard, text extraction) but no test coverage. The hook and component tests are solid, but this integration point is untested.

---

None of these are blockers for the current MVP scope, but the code-blocks-read-aloud issue would be worth fixing before merge — it'll be immediately noticeable in use.

🤖 Generated with Claude Code