native: visionOS voice input via Deepgram relay

[itomek]

Part of #125. Blocked by the visionOS shell.

Problem

Typing in a headset is slow. Claudia already has Deepgram voice plumbing (voice WS relay + POST /api/tts) used by the web clients — port the input half to visionOS.

Outcome

The composer ornament's mic orb streams speech through the backend's Deepgram relay and inserts the transcription into the composer (push-to-talk).

Acceptance criteria

• Push-to-talk: hold/tap mic → live partial transcription visible → release finalizes into the composer
• Microphone permission flow handled; no-Deepgram-key state degrades to a disabled mic with explanation
• Works on simulator with a mic-enabled Mac; device pass documented if hardware available
• Audio session configured to not disrupt other audio

Scope & expectations

• Where: visionOS target + ClaudiaKit voice client (reuse backend relay endpoints as-is)
• Out of scope: TTS playback of summaries, always-listening mode

How to verify

UI — recording of a voice-dictated prompt creating/steering a task.