fix(serve): prevent echo-trigger loops that break demo TTS

[crrow]

Summary

• fix demo regressions where speech-recognition echo/noise could repeatedly spawn new turns and close /ws/tts before streaming started
• replace immediate onspeechstart hard interrupt with final-text-confirmed interruption
• add assistant-echo guards to reject recognized fragments that mirror assistant output
• gate recognition auto-restart with an explicit run flag to avoid stop -> onend -> restart races while entering THINKING

mlx-live Comparison

mlx-live uses backend VAD-triggered interrupt and explicit interrupt events to stop playback, instead of immediate browser-side speechstart interruption:

• interrupt_active_response in main.py
• frontend handles {"type":"interrupt"} by clearing playback in web/index.html

This PR aligns our demo behavior with that intent by making interruption confirmation stricter and less echo-prone.

Validation

• cargo test serve::tests -- --nocapture
• Playwright deep baseline (deep-3477.js): binary=1 decoded=1 played=1
• Playwright barge-in + think filter (barge-think-3477-v2.js):
  • LLM_CALLS=2
  • THINK_IN_WS_SEND=false
  • final status returns to listening…