Full-duplex voice mode for DeepSeek Harness: streamed ASR → LLM → TTS with barge-in. Local whisper transcription, Edge TTS playback, zero API key.