fix(audio): preserve speech through echo cancellation

[kvnloo]

Summary

• suppress speaker feedback before it reaches server VAD
• use WebRTC speech detection for fallback barge-in gating
• route default Linux audio through PulseAudio WebRTC echo cancellation/noise suppression
• bypass the fallback amplitude gate once upstream AEC is active, preventing quiet words and numeric ranges from being clipped

Root cause

PulseAudio had already removed playback echo, but the application applied a second amplitude/VAD filter and discarded full 100 ms microphone blocks during assistant playback.

Verification

• pytest -q tests/test_audio.py — 18 passed
• ruff check . — passed
• regression proves quiet input survives assistant playback when PulseAudio AEC is active

[kvnloo]

Review wave complete. Fixed fallback speech-gate clipping, Pulse route fallback, environment ownership/restoration, concurrent route naming/lifetimes, and WebRTC VAD availability reporting in 934cafc. Verification: 24 audio tests passed; Ruff clean.