feat(voice): provider-agnostic TTS seam + audio-reactive renderer

[reshimu]

Summary

• Provider-agnostic voice seam (src/voice/): a VoiceAdapter interface with an ElevenLabs adapter (proxied through server-side api/voice/*.js so the key never reaches the bundle) and a zero-cost Web Speech fallback, selected via VITE_VOICE_PROVIDER.
• VoiceViz — audio-reactive canvas renderer driven by the pre-written voice-drivers.js (smoothed RMS/low-band/spectral-centroid scalars), accepting either a live AnalyserNode or a pre-computed curve.
• VoicePanel — the connection surface: provider name, key present/absent, health() result, and a Test voice button wired into VoiceViz.
• Spec at docs/specs/voice-viz.md, approved before implementation.

Test plan

• pnpm typecheck — clean
• pnpm lint — clean (--max-warnings=0)
• pnpm test — 19 files / 87 tests passing (voice-drivers TDD suite, adapter tests with mocked fetch/speechSynthesis, VoiceViz smoke test)
• pnpm build — demo + library both build; public library bundle size unchanged (voice code isn't exported publicly)
• pnpm check:bundle-size / test:consumer / test:react-consumer — all pass
• Two-stage review: spec compliance PASS, code quality PASS (two real bugs found and fixed — a stuck-disabled Test voice button in the default config, and an AudioContext/blob-URL leak on repeated clicks)
• Verified no secret staged (git diff --cached grep, .env.example has empty values, .gitignore covers .env*)

🤖 Generated with Claude Code