feat(voice): ElevenLabs ASR + line-style composer icons

[oratis]

Voice ASR

Transcription now prefers ElevenLabs Scribe as the primary provider, falling back to OpenAI Whisper:

• Provider order: ELEVENLABS_API_KEY → POST https://api.elevenlabs.io/v1/speech-to-text (multipart file + model_id, xi-api-key header).
• Falls back to OpenAI Whisper when no ElevenLabs key is present.
• The signature is unchanged; the no-key error names both providers.
• New src/voice/transcribe.test.ts adds 3 tests (mock fetch + temp file).

Composer icons

• Composer #plusBtn (＋) and #recordBtn (🎙) glyph/emoji replaced with line-style SVGs matching the .fbtn function bar.
• CSS: color --fg-3 → --fg-2, dropped glyph font-size, added #plusBtn svg,#recordBtn svg{width:19px;height:19px;display:block}.
• HTML snapshot length 145995 → 146593 (sha256 recomputed).

Verification

• npm run typecheck → clean
• npm test → 889 passing (1 skip)

🤖 Generated with Claude Code

[oratis]

Reviewed + rebased onto latest main (v0.14.0). Dropped 3 commits that had independently landed on main (providers-cache #210, docs-publish #212, and the /chat empty+abort logic already merged via #206). Resolved the snapshot conflict by recomputing against the merged lisa-html.ts (QR-pairing #171 + these icon changes): EXPECTED_LENGTH 151331, sha256 72098d77…. typecheck clean; voice + snapshot tests green; CI green. Code review: transcribe.ts cleanly splits per-provider helpers, no-key error names both, non-2xx + missing-text guarded; HTML gained aria-labels. LGTM.