feat(voice): first-class voice input and local TTS across web, desktop, and mobile (#2018)

Complete rebuild of voice input on a server-authoritative streaming
architecture, replacing the legacy Web Speech / whole-blob / WASM engines
and the dead voice-agent layer (~4k lines removed).

Speech-to-text (dictation):
- Client streams 16 kHz mono PCM16 chunks over /api/dictation/ws with
  seq/ack ordering; buffered audio is retained and replayed on reconnect
- Server transcribes and streams live partial transcripts back;
  segments auto-commit every ~15s with silence suppression and adaptive
  finalization timeouts
- Local provider (default, zero config): sherpa-onnx models in a forked
  worker process — auto-download with progress, staged extraction with
  verification, corrupt-model auto-recovery, idle shutdown after 5 min
- Model catalog with settings picker (accuracy/speed ratings, sizes,
  download/delete): Parakeet TDT v2 (English) and v3 (25 European
  languages, auto-detected), Whisper base and tiny (multilingual, light)
- OpenAI-compatible provider for any Whisper endpoint
- Composer overlay with live transcript, volume meter, timer, and
  cancel / insert / insert-and-send actions; failed transcriptions keep
  their audio for retry or accepting the partial text as-is
- Configurable keyboard shortcut (default mod+alt+v) toggles dictation;
  Enter confirms and Escape cancels while recording
- Overlay is pixel-aligned with the composer (measured footer height,
  matching paddings/typography/gaps) — no layout shift when toggling

Text-to-speech:
- Local Kokoro provider (English, 11 voices) synthesized in the same
  worker via /api/dictation/tts/speak, managed by the shared model
  pipeline; sentence-pipelined playback keeps time-to-first-audio at
  ~1 sentence regardless of message length, and stop cancels in-flight
  synthesis
- Sanitizer keeps inline-code content (strips backticks only), reads
  interword slashes aloud, and removes only absolute file paths

Settings:
- Voice page unified: a single read-aloud toggle owns all playback
  options (the confusing "Enable Voice Mode" is gone); a new "Enable
  voice input" toggle (default on, persisted to settings.json) hides
  the composer mic entirely when disabled

Mobile and transport:
- iOS/Android microphone permissions added (dictation was previously
  impossible on mobile)
- Fixed Android WebSocket upgrades: the Capacitor WebView origin
  (https://localhost) was missing from the packaged-client allowlist,
  403-ing every WS connection — root cause of the old mobile SSE lock,
  which is now removed for all transports

Security and conventions:
- All HTTP routes sit behind the global /api auth gate; the WS upgrade
  explicitly validates the UI session and origin, with oc_url_token
  narrowly allowlisted and covered by tests; the dictation socket mints
  a fresh URL token before connecting
- Routes register before the generic OpenCode proxy; the client goes
  through runtimeFetch/getRuntimeUrlResolver, and runtime switches
  reset the dictation socket
- VS Code deliberately reports dictation as unavailable (no server
  process in that runtime)

CI: workflow Node bumped 20 -> 22 to match the repo engines and fix
better-sqlite3 installs broken by node-gyp@latest on Node 20.

New dependency: sherpa-onnx-node (prebuilt N-API; macOS/Linux x64+arm64,
Windows x64 — Windows-on-ARM falls back to the OpenAI-compatible provider)

packages/web/server/lib/dictation/DOCUMENTATION.md

@@ -0,0 +1,63 @@
+# Dictation module
+
+Server-authoritative streaming speech-to-text for the chat composer, plus
+local text-to-speech. The client streams 16 kHz mono PCM16 chunks (base64)
+over a WebSocket; the server runs the transcription and streams live partial
+transcripts back.
+
+Local TTS (Kokoro via sherpa-onnx OfflineTts) runs in the same worker process
+and is exposed as `POST /api/dictation/tts/speak` (JSON `{text, speakerId?,
+speed?, model?}` → WAV bytes; 503 with `reasonCode` while the model is
+downloading). TTS models live in the same catalog/downloader as STT models
+(`local/model-catalog.js` `LOCAL_TTS_MODEL_CATALOG`) and are managed by the
+same status/download/delete routes.
+
+## Ownership
+
+- `runtime.js` — registers `GET /api/dictation/status`,
+  `POST /api/dictation/models/:modelId/download`, and the
+  `/api/dictation/ws` WebSocket endpoint (auth-gated the same way as the
+  terminal WS: UI session token or `oc_url_token`, plus origin check).
+  Created from the startup pipeline (`startup-pipeline-runtime.js`) before
+  the generic OpenCode proxy so routes are not shadowed.
+- `stream-manager.js` — `DictationStreamManager`, one per WS connection.
+  Chunk reordering by `seq` + ack, resampling to the provider rate,
+  auto-commit every ~15 s of audio, silence suppression by PCM peak,
+  partial-transcript concatenation, adaptive finalization timeout.
+- `service.js` — provider resolution and readiness. Providers:
+  - `local` (default): sherpa-onnx Parakeet TDT in a forked worker process.
+    Models auto-download in the background on first use; while missing, the
+    stream fails with `reasonCode: 'model_download_in_progress'` and the
+    status route reports per-model install/download state.
+  - `openai-compatible`: buffered per-segment transcription against any
+    OpenAI-compatible `/v1/audio/transcriptions` endpoint
+    (`openai-compatible-session.js`, reuses `../tts/stt.js`).
+- `local/` — worker process + client (IPC, idle shutdown TTL), sherpa
+  recognizer engine and realtime session (throttled re-decode for partials),
+  model catalog and downloader. The native `sherpa-onnx-node` addon is only
+  ever loaded inside the worker process.
+- `audio.js` — PCM16 helpers: format parsing, peak, WAV wrapping, streaming
+  linear resampler.
+
+## WebSocket protocol (JSON text frames)
+
+Client → server: `start {dictationId, format, options}`,
+`chunk {dictationId, seq, audio}`, `finish {dictationId, finalSeq}`,
+`cancel {dictationId}`, `ping`.
+
+Server → client: `ready`, `ack {ackSeq}`, `partial {text}`,
+`finish_accepted {timeoutMs}`, `final {text}`,
+`error {error, retryable, reasonCode?}`, `pong`.
+
+`options` in `start` carries the client-selected provider config:
+`{ provider: 'local' | 'openai-compatible', language?, localModel?,
+openaiCompatible?: { baseUrl, model, apiKey } }`.
+
+## Invariants
+
+- Never load `sherpa-onnx-node` in the main server process.
+- The stream manager acks only the highest contiguous seq; the client is
+  expected to retain unacked segments for retry/replay.
+- Silence-only segments (peak < 300) are cleared, never committed, so
+  Whisper-style providers do not hallucinate on silence.
+- Model files live under `~/.config/openchamber/speech-models`.