Add a WebRTC realtime voice provider on Android (full-duplex speech models)

[s546126]

Summary

Adds a WebRTC-based realtime voice path on Android, opening VisionClaw up to full-duplex speech-to-speech models beyond Gemini Live.

• voice/RealtimeVoiceService — a provider-neutral interface extracted mechanically from GeminiLiveService's existing public surface (connect/disconnect, audio, video frames, tool responses, the eight callbacks). GeminiLiveService implements it unchanged in behavior.
• voice/openai/OpenAIRealtimeService — an OpenAI Realtime API backend over WebRTC, using the documented ephemeral-key flow (POST /v1/realtime/client_secrets, then SDP exchange with POST /v1/realtime/calls). Model/tool events flow over the oai-events data channel. Reuses the stream-webrtc-android dependency already in the project.
• Settings — a Voice Provider toggle (Gemini Live remains the default) plus an OpenAI API key field. Stored in SharedPreferences only; no Secrets.kt changes, so existing local setups keep compiling.

Why WebRTC for voice

Full-duplex conversation lives or dies on barge-in, and that's where a raw PCM-over-WebSocket pump struggles: speaker audio leaks back into the mic, and the model interrupts itself. WebRTC brings echo cancellation, jitter buffering, packet-loss concealment, and Opus for free — the peer connection owns the mic and speaker end to end (managesOwnAudio = true), so AudioManager's PCM pipeline stays idle on this path. A side benefit: this sidesteps the known Gemini-Live-vs-WebRTC audio device conflict for providers on this path.

The interface also gives other full-duplex voice backends a contained place to land as additional RealtimeVoiceService implementations, without touching the session/tool-routing layers again.

Design notes

• Tool calls: ToolCallRouter now hands back (callId, name, result) and each provider shapes its own wire format (Gemini toolResponse JSON vs OpenAI function_call_output + response.create). The Gemini JSON building moved verbatim from the router into GeminiLiveService.
• Vision: Gemini takes a continuous ~1fps frame stream; pushing that through Realtime conversation items would flood context. The OpenAI path instead caches the latest camera frame and attaches it as an input_image item once per user turn, on input_audio_buffer.speech_started.
• Provider switch takes effect on the next session — the service is re-created in startSession().
• Teardown from WebRTC observer callbacks hops to a separate thread (PeerConnection.close() blocks on observer callbacks returning).

Testing

• Compile-verified: ./gradlew assembleDebug passes (DAT SDK 0.4.0 via GitHub Packages).
• Not yet exercised against live glasses hardware; happy to iterate on review feedback.
• iOS is untouched; the same interface extraction would be the follow-up there.

🤖 Generated with Claude Code