fix(audio): prevent main-thread hang after Bluetooth device route change

[thedavidweng]

Address #77

Summary

Fix the app-freeze bug where Koe hangs with 30%+ CPU after a Bluetooth headphone / audio device route change. The hotkey triggers but recording never starts; the process stays alive but unresponsive until force-quit.

Root Cause

Three cascading failures after a BT device route change:

1. AudioUnitSetProperty fails with 'nope' (1852797029) when setting the input device on the new AVAudioEngine. The previous code logged this but continued with the corrupted engine whose IO unit was in an inconsistent state.

2. inputNode.outputFormatForBus:0 returns 0 ch, 0 Hz on the broken engine. No validation existed, so an invalid format was passed to installTapOnBus: and startAndReturnError:.

3. AVAudioEngine.start() blocks indefinitely — CoreAudio's HALC_ProxyIOContext::StartAndWaitForState returns error 35 (EAGAIN) and never completes. Since this runs synchronously on the main thread, the entire app freezes.

[Koe] Audio device list changed
HALC_ProxyIOContext::_StartIO(): Start failed - StartAndWaitForState returned error 35
AVAudioIONodeImpl.mm:1097 Error setting device on iounit, err = 1852797029
Input render format:  0 ch,      0 Hz

Fix

Three defensive layers, any one of which prevents the hang:

1. Discard corrupted engine on AudioUnitSetProperty failure

When setting the input device fails, the IO unit is in an unusable state. Instead of proceeding, create a fresh engine that falls back to the system default input device.

if (osStatus != noErr) {
    self.audioEngine = [[AVAudioEngine alloc] init];
    inputNode = self.audioEngine.inputNode;
}

2. Validate inputNode format before proceeding

Return NO early if channelCount == 0 or sampleRate <= 0:

if (hardwareFormat.channelCount == 0 || hardwareFormat.sampleRate <= 0) {
    self.audioEngine = nil;
    return NO;
}

3. Move audioEngine.start() off the main thread with a 3s timeout

Use dispatch_semaphore to bound the blocking start() call. If CoreAudio's HAL proxy hangs, the main thread is released after 3 seconds, the error path triggers, and the user can retry immediately.

dispatch_async(dispatch_get_global_queue(QOS_CLASS_USER_INITIATED, 0), ^{
    startOK = [engine startAndReturnError:&bgError];
    dispatch_semaphore_signal(sem);
});
long timedOut = dispatch_semaphore_wait(sem, dispatch_time(DISPATCH_TIME_NOW, 3s));
if (timedOut) { /* abort cleanly */ }

Also

• Extract duplicated hold/tap audio-start code into startAudioCaptureWithRetry with a single 500ms retry for transient post-route-change failures
• Release audioEngine = nil on all failure paths and in stopCapture to prevent stale engine refs

Behavior

Scenario	Before	After
BT headphone route change during idle	Next hotkey press → app freezes forever	Falls back to system default mic, or retries in 500ms, or shows error and recovers in 2s
BT headphone disconnects mid-recording	Works (existing device-list listener)	Works (unchanged)
Normal operation (no device change)	Works	Works (no timeout overhead — start() returns in <100ms)
AudioUnitSetProperty returns 'nope'	Continues with broken engine → hang	Discards engine, creates fresh one with default device
inputNode reports 0 ch, 0 Hz	Passes to installTap → crash/hang	Returns NO immediately

Testing

• All Rust tests pass (cargo test --no-default-features — 56 tests)
• Modified .m files verified for balanced braces/brackets/parens
• cargo fmt --check — pass

[missuo]

@thedavidweng Please resolve the conflicts.

[thedavidweng]

@cursoragent Resolve conflicts

[missuo]

@cursoragent Resolve conflicts

It seems that cursor ignores you.

[thedavidweng]

@cursoragent Resolve conflicts@cursoragent 解决冲突

It seems that cursor ignores you.看起来光标会忽略你。

你大爷的