voice: make passive ptt_start the canonical hands-free narration fix

[accnops]

Problem

In hands-free mode the client opens a passive barge-in mic that keeps streaming while it waits for the user to actually speak. Auto-listen does the same. Those ptt_start events made the backend latch the session as "user is speaking" for the whole open-mic window, and the window ends silently (the mic is aborted with no ptt_end), so the latch never cleared. Client-driven narration requested during that window was NACK'd busy and stranded. A few reconnect and unblock paths also lost or failed to replay narrations that were deferred while the session was busy.

#326378 fixed the same symptom from the frontend by force-ending the open turn (synthesizing a ptt_end) before narrating. That works, but it overloads ptt_end to mean "stop for playback" and can cut a live barge-in short.

Solution

Address the root cause instead of forcing an end. The client tags every silent open-mic turn (barge-in and auto-listen) with passive: true on ptt_start; the backend sets user_is_speaking = not passive, so a passive turn never latches. Deliberate presses stay non-passive and still latch, which correctly defers narration while the user is actually speaking.

Because this makes the passive flag the canonical fix, it reverts the #326378 frontend workaround (now on main): _userSpeechActive, _isActivelyDictating(), _finishPtt('immediate'), the _prepareForPlayback(endOpenTurn) parameter, canRequestNarration, and the endPassiveTurnFirst block in _narrate.

Supporting changes on the branch:

• Adopt the server session id and clear the resuming flag on (re)init.
• Replay deferred narrations on reconnect.
• Retry the specific narration id named by narration_unblocked instead of the latest.
• Client-side telemetry when a narration is deferred or dropped.
• Preserve a held deliberate press in _prepareForPlayback: a non-passive press latched user_is_speaking, so its narration was deferred; aborting it here would send no ptt_end and strand the latch. Its natural release sends ptt_end, clears the guard, and drives the narration_unblocked retry. Passive open-mic turns (which never latched) are still torn down to free the mic for narration audio. _prepareForPlayback now returns whether the slot is ready.
• Don't play or double-speak buffered narration over a held deliberate press. _flushDeferredResponse peeks the buffered response, prepares the slot, and only deletes/enqueues it when the slot is actually ready; a preserved press keeps the buffer intact and reports retained plus the buffered reply's transcript. _activateShownSession then skips a fresh narration only for the SAME reply that buffer will replay on release (matched by normalized transcript). An unrelated confirmation or a newer, different response for the session still narrates, because a stale buffered response can outlive its turn (a thinking transition clears _deferredNarrations, not _deferredResponses). The buffer plays on release via the periodic safety-net; the pending indicator stays until then, so nothing is lost.
• Clear the passive flag when a barge-in listen is promoted to a real press, so a promoted press latches locally and its narration is preserved rather than torn down.

Tests

voiceSessionController.test.ts and voiceClientService.test.ts:

• Auto-listen opens a passive mic turn; a deliberate press opens a non-passive one.
• A held deliberate press awaiting narration is preserved (no abort, stays held).
• A held deliberate press keeps buffered narration deferred instead of playing over the press.
• A buffered reply retained under a held press is not re-narrated (no duplicate on release).
• A different pending reply is still narrated when an unrelated buffer is retained under a held press.
• Promoting a passive barge-in listen to a real press clears the passive flag so playback preserves the press.
• A passive open-mic turn is torn down for playback since it never latched.
• Deferred-narration replay on reconnect, unblock-by-id retry, and the deferred/dropped telemetry paths.

Full voiceClient suites pass (controller 19, client service 17, glow 3). ESLint clean; typecheck clean (only the 8 pre-existing native-module errors unrelated to this change).

Risks

• Depends on the backend understanding passive (yolo#56284). This must reach GA before this lands. If it merges first, the backend ignores passive and, with #326378 reverted, hands-free narration would strand again. Kept as a draft for that reason.
• Reverts merged frontend code from #326378; anyone relying on the forced-ptt_end behavior would see it change to the passive model.

Non-goals

• No change to how narration text is phrased or how audio is echoed.
• No new dependencies, settings, or public API surface.
• The narrow async cold-mic acquisition race is pre-existing and out of scope; with VAD it self-corrects via the backend speech edge.
• Making a promoted barge-in press visible on the wire. Clearing the passive flag on promotion is local state only; the turn's ptt_start was already sent as passive: true, so the backend still treats it as passive and does not latch. With VAD, real speech during the press cancels any narration via the speech edge. Without VAD there is no speech edge, so a narration requested during that brief promoted-press window can be accepted and play over the user. A wire-level promotion that latches the backend for the same turn is deferred (it needs a protocol change and carries regression risk on the working deliberate-press path).
• Clearing a session's pending-reply indicator after the periodic safety-net (rather than a focus) replays its buffered audio. The safety-net flush currently ignores the flush result, so the indicator can linger and a later refocus can re-narrate the reply. This is pre-existing (unchanged by this PR) and tracked as a follow-up.

[Copilot]

Pull request overview

Improves voice narration recovery during hands-free listening and reconnects.

Changes:

• Marks passive barge-in PTT sessions.
• Restores session and deferred narration state after reconnects.
• Adds narration defer/drop telemetry and tests.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
voiceClientService.test.ts	Tests PTT serialization and session recovery.
common/voiceClientService.ts	Extends the PTT API with passive.
voiceTelemetry.ts	Defines narration lifecycle telemetry.
voiceSessionController.ts	Replays and revalidates deferred narrations.
browser/voiceClientService.ts	Serializes passive PTT and adopts server session IDs.
micCaptureService.ts	Preserves passive state through microphone acquisition.