Full-duplex voice plugin for DeepSeek Harness

Key features

• The llm/stream tap is losslessevery chunk is yielded unchanged, the segmenter only observes.
• The model stream is never blocked by synthesis.
• ASR runs host-side with sherpa-onnx (Apache-2.0) running SenseVoice — the CN-native speech model that outperforms whisper on Chinese: native simplified output, punctuation, inverse text normalization (ITN) and 50+ language auto-detection.
• The browser only records and posts raw little-endian f32 PCM.