fix(interruption): don't drop committed audio at the transport send boundary

.changeset/adaptive-interruption-send-gate.md

@@ -0,0 +1,11 @@
+---
+'@livekit/agents': patch
+---
+
+Adaptive interruption: stop dropping audio at the send boundary, and make the verdict diagnosable.
+
+The WebSocket transport re-checked `overlapSpeechStarted` immediately before writing to the socket, after awaiting any in-flight reconnect. Because `overlap-speech-ended`, `agent-speech-ended` and `bargein_detected` can all clear that flag inside the await window, audio the pipeline had already committed to sending could be discarded and never counted in `numRequests`. The buffering stage upstream is the only place that decides whether a slice belongs to an overlap, which matches the Python implementation, whose send task gates on nothing.
+
+Slices now carry the overlap they were cut for, so a send that lands after that overlap closed is no longer counted against whichever overlap happens to be open when the socket accepts it. Without this, a later overlap could report a `numRequests` it never made.
+
+An overlap that ends without any usable inference result now logs at `warn` (previously `debug`) with the overlap duration, `numRequests`, buffered samples and agent-speech state, so a fallback backchannel verdict is no longer indistinguishable from a genuine low-probability one. Every verdict also logs `probability`, `isInterruption` and `numRequests` at `debug`. `OverlappingSpeechEvent` is now exported by name for typing `overlapping_speech` handlers.

.changeset/fix-recorder-capture-finish-race.md

@@ -0,0 +1,15 @@
+---
+'@livekit/agents': patch
+---
+
+Fix a deadlock where a recorder-wrapped audio output could leave `waitForPlayout` stranded when
+an interrupt arrived before the recorder had registered its segment. `RecorderAudioOutput` now
+registers its own segment before forwarding a frame downstream, and attributes each playback
+finish to the segment it belongs to instead of relying on a global counter.
+
+A recorded segment is also timestamped when it opens rather than when the wrapped output accepts
+its first frame, so a finish that lands while that frame is parked no longer clamps the segment's
+playback position to zero and drop the audio the sink reported as played. And a segment whose
+downstream capture throws now releases the capture latch on both the recorder and the wrapped
+output, so a caller that retries after a transient rejection is no longer rejected forever with
+`recorder capture has no active segment`.