Voice AI Latency Budget: Fast and Slow Models in Parallel

Map endpointing, transcription, model branches, TTS, and playout before optimizing the wrong stage.