fix(parakeet): config-change restart loop killed the mic in v1.0.12 (hotfix)

[initcore0]

The v1.0.12 regression

Symptom (user report): mic completely dead on every dictation in v1.0.12; downgrading to v1.0.11 fixes it. v1.0.12's only functional change is #224.

Root cause: #224's AVAudioEngineConfigurationChange handler restarts capture on every notification. But the notification also fires for io-unit changes the handler's own teardown+rebuild causes — and on the reporting machine a capture start itself posts one. So the first notification at session start enters a self-sustaining loop:

notification → teardown+rebuild → engine.start() posts notification → teardown+rebuild → …

~4 restarts/second for the whole session; capture never lives long enough to feed the decoder; every session times out with no speech.

Reproduced live against the released artifacts on the affected machine (warm model, cold-load noise excluded):

• v1.0.12 + pinned Shure MV7+: 35 restarts in a 10s session, no transcript.
• v1.0.12 + BlackHole fixture harness: 65 restarts in 15s, no speech within the time limit; identical harness on v1.0.11 transcribes correctly.
• stderr shows the storm: [Parakeet] input device configuration changed — restarting capture every ~240ms.

Fix

The restart is no longer reflexive:

• A notification now schedules a settle-check (settleDelay 0.3s) instead of restarting inline; further notifications in the window coalesce into the pending check.
• The check rebuilds only when the engine genuinely stopped (AVAudioEngine.isRunning == false at settle time) — the real device-loss case the observer exists for (fix(parakeet): restart capture on mid-session input-device configuration change #224's goal). A notification that leaves the engine running is a transient renegotiation and is ignored; restarting a running engine is exactly what fed the loop.
• Rebuilds are budgeted (maxRestartsPerSession = 3, reset per capture session). A device that still stops capture after that fails the session loudly (gaveUpMessage) instead of looping silently — fix(parakeet): restart capture on mid-session input-device configuration change #224's word-loss guarantee is preserved, without the loop.

Decision matrix lives in core CaptureConfigChangePolicy (tested); the engine holds only the observer, the coalescing flag, and the rebuild.

Verification

• This build, same BlackHole harness, warm model: "The quick brown fox jumps over the lazy dog", exit 0, zero capture restarts, across three consecutive sessions.
• swift test: 1908 green (new policy matrix cases: running→ignore, stopped→restart, budget→giveUp, stale fences).
• ./build.sh clean.

Notes

• The genuine mid-session device-unplug path (fix(parakeet): restart capture on mid-session input-device configuration change #224's scenario) still needs a live unplug/replug pass — the policy now requires isRunning == false, which a real device loss produces, but that's asserted by reasoning + tests, not by a live unplug in this run.
• fix(engines): restart capture on mid-session input-device config change (WhisperKit, Apple Speech, SpeechAnalyzer) #226 (config-change observers for WhisperKit/Apple Speech/SpeechAnalyzer, open) copies fix(parakeet): restart capture on mid-session input-device configuration change #224's reflexive-restart pattern and needs the same settle-check treatment before merging.
• Recommend cutting a release with this fix promptly — v1.0.12 is unusable for Parakeet (default engine) users on affected hardware.

🤖 Generated with Claude Code