[johnmatthewtennant]

Summary

Berd Call kept holding the agent's reply after the user had finished, and showed the user's words only once Apple finalized them. This collects the release-safe fixes from the call learning work, without its capture, replay or tuning infrastructure.

Stuck speech holds. The hold that keeps an agent from speaking over the user is released by Apple's finals, but Apple emits one final for a sentence that pauses split into several VAD turns. Resolving one turn per final left the rest outstanding, so in recent calls 55 of 57 holds ran the full 5 s timeout even though Apple had finished 0 to 4 s after the user stopped. Expiry also cancelled the recognizer, throwing away a transcript still on its way, and the hold lifted at the timeout while Apple was still transcribing, so a reply started and was then cut off by the late transcript.

• Expiry releases only the hold. Recognition, its queued audio and the recognizer generation continue, so a late final still reaches the conversation.
• Each final carries the end of its audio through the Swift bridge and resolves exactly the ended turns it covers. A final that arrives during the silence window, before its turn ends, resolves that turn when it ends. A final for an earlier phrase of a turn still being spoken resolves nothing.
• The bridge reports Apple's interim results. While Apple's latest result is interim words, the hold lasts up to 10 s beyond the 5 s timeout, until a final follows. Noise with no interim words keeps 5 s.

Interim words. Conventional macOS sessions emit user_interim (--stream carries it as an ephemeral 0\tuser_interim row) with Apple's interim text, replaced by each update and cleared on its final, mute, input reset or restart. It is never stored, never takes a token and never interrupts output.

One VAD threshold. Listening and interruption during agent playback use the same speech probability, 0.9, instead of separate thresholds.

Playback and cancel fixes. A cancel wake could be lost between a playback worker's flag check and its wait, idle live playback did not wake on cancellation, and handoff terminal failures lost their order.

What came from where

Commit	Source
wake idle live playback on cancellation	80b97e4c2 in #426 (cherry-pick)
cover a cancel wake racing the flag check (test)	f6097e517 in #426 (cherry-pick)
order cancel wakes after the flag check	14943c91e in #426 (cherry-pick)
preserve ordered handoff terminal failures	6b5afdf00 in #426 (cherry-pick)
tolerate audio replies after child shutdown (test)	46fdf603b in #426 (cherry-pick)
validate partial audio EOF after intentional disconnect (test)	71214694e in #426 (cherry-pick, conflict resolved by dropping the capture-only replay test)
expire speech holds without cancelling recognition	2eda9918c in #426 (ported)
release the speech hold when a final covers its turns	43efb0e00 in #426 (ported)
keep the speech hold while Apple still has words to finalize	09ee43fd9, e173655ac, 4d2f03488 in #426 (ported, final combined behavior)
stream the user's words while they are transcribed	7b28c3ecd in #428 (ported)
ignore interim words in native voice conversations	e347e12a6 in #428 (cherry-pick)

The ports differ from #426 only where #426 depended on infrastructure this branch leaves out:

• The pending-utterance timeout is a fixed 5 s constant and the interim grace a fixed 10 s, not DetectionSettings fields, CLI flags or a protocol bump.
• Capture hooks and the capture-only heard and settled records are gone, along with the detection-evidence handling in finals (result_is_current still gates hold accounting).
• A turn's end is the pushed audio minus main's 1.2 s silence flush window, where berd-call: capture calls and replay speech detection profiles #426 subtracted its configurable release_ms.

Left in #426

Capture and replay, VAD debug cues, failure fixtures, speech-profile evaluation, and the learning and tuning work (dynamic timeouts, thresholds and release).

Related issue

Follows #426 and #428. None found otherwise.

Testing

• Real Apple recognition: the ignored on-device test mac_speech::tests::final_end_is_measured_on_the_pushed_audio_timeline passes, so a final's end is on the same pushed-audio timeline the VAD turns use. That is what makes the coverage rule sound.
• Hold release, red/green: mac_final_before_its_turn_ends_releases_the_hold_when_it_does fails before the coverage fix (the turn stays pending until the 5 s timeout) and passes after it. Companion tests cover a final resolving every turn its audio covers, an earlier-phrase final keeping the hold, the hold waiting for interim words and returning to 5 s once a final follows, and late finals surviving hold expiry.
• Interim words: tests cover the stream through mute, reset and restart, and the interim token matching the next final.
• Playback cancellation: the cherry-picked session_protocol regressions run the real berd-call child against a scripted audio host.

Not yet exercised on a live call: the end-to-end hold timing and interim captions with a real microphone.