feat(voice): dictate through the server's Claude Code login, no API key

The mic button previously needed a Deepgram API key, or fell back to the
browser's Web Speech engine. It can now transcribe through the same
speech-to-text service Claude Code's own /voice mode uses, so anyone signed
in to Claude Code on the server gets dictation with no third-party account.

Claude Code's voice mode cannot be driven directly: it opens the HOST's
microphone (sox/arecord), and the CLI runs in a headless tmux pane while the
human is in a browser somewhere else. So capture stays in the browser and only
the transcription backend is borrowed.

Audio goes browser -> Codeman -> Anthropic. The OAuth token never reaches the
page: the browser sends PCM16 (16 kHz mono, produced by an AudioWorklet since
MediaRecorder cannot emit raw PCM) and receives text.

- GET /api/voice/status reports readiness and never the token
- GET /ws/voice/stream relays one dictation, with the same Host/Origin upgrade
  guard as the terminal socket, plus caps on concurrency, stream length and
  frame size
- credentials are read-only: Codeman never refreshes them, since a refresh
  rotates the refresh token and could sign the user out of their own CLI
- claudeVoiceEnabled (synced, default OFF) gates the whole server side
- voiceSettings.provider picks auto/claude/deepgram/webspeech; auto prefers
  Claude, then a configured Deepgram key, then the browser

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

src/web/public/voice-pcm-worklet.js

@@ -0,0 +1,57 @@
+/**
+ * @fileoverview AudioWorklet that turns microphone audio into the PCM frames the
+ * Claude voice endpoint expects.
+ *
+ * The endpoint is opened as `encoding=linear16, sample_rate=16000, channels=1`,
+ * i.e. raw signed 16-bit little-endian mono. MediaRecorder cannot produce that
+ * (it only emits container formats — webm/opus, mp4), which is why the Deepgram
+ * path's capture code cannot be reused here: Deepgram sniffs the container,
+ * Anthropic's endpoint does not.
+ *
+ * Sample rate is handled by the AudioContext, constructed at 16000 Hz so the
+ * browser resamples the mic for us. This processor only converts Float32 [-1,1]
+ * to Int16 and batches, because a raw 128-sample render quantum is a ~4 ms
+ * WebSocket frame — 250 frames a second of pure overhead.
+ *
+ * Loaded via `audioWorklet.addModule()` from voice-input.js. Runs on the audio
+ * thread: no DOM, no globals from the page.
+ *
+ * ⚠️ Edit this file and voice-input.js together. Static assets are served
+ * `immutable` for a year and this one is fetched from JS, so it inherits its
+ * cache-bust token from voice-input.js's script tag (see `_workletUrl()`); a
+ * change here alone would keep serving the old copy to every returning browser.
+ */
+
+/** ~256 ms at 16 kHz. Big enough to keep frame overhead down, small enough that interim transcripts stay live. */
+const FRAME_SAMPLES = 4096;
+
+class PcmFrameProcessor extends AudioWorkletProcessor {
+  constructor() {
+    super();
+    this._buffer = new Int16Array(FRAME_SAMPLES);
+    this._offset = 0;
+  }
+
+  process(inputs) {
+    const channel = inputs[0]?.[0];
+    // No input yet (mic still warming) — keep the processor alive.
+    if (!channel) return true;
+
+    for (let i = 0; i < channel.length; i++) {
+      // Clamp before scaling: values slightly outside [-1,1] are legal in Web Audio
+      // and would wrap around to the opposite sign as Int16, which sounds like a click.
+      const sample = Math.max(-1, Math.min(1, channel[i]));
+      this._buffer[this._offset++] = sample < 0 ? sample * 0x8000 : sample * 0x7fff;
+
+      if (this._offset === FRAME_SAMPLES) {
+        // Transfer a copy: the worklet keeps reusing its own buffer.
+        const frame = new Int16Array(this._buffer);
+        this.port.postMessage(frame.buffer, [frame.buffer]);
+        this._offset = 0;
+      }
+    }
+    return true;
+  }
+}
+
+registerProcessor('pcm-frame-processor', PcmFrameProcessor);