Per-turn view with end-to-end pipeline latency (STT → agent → TTS). Transcript/response prefer voice API fields; chat history is a fallback.