Listen-loop capture now only commits when the caller actually speaks:
phase A waits for a sustained voice onset (3 frames above threshold,
rejecting the PA-mute click), phase B records until silence. Empty
captures are no longer posted, so the NPC never replies to silence.
Also: DC-blocking high-pass on the captured mono, VAD thresholds tuned
to the quiet SLIC handset mic (onset ~1.4%, silence ~0.6%), keep the
PA on during capture (muting it collapsed the mic), and add a
/debug/miccap diagnostic endpoint (raw fixed-duration mic capture).
clement