fix: prevent audio truncation and eliminate recording start delay

[jatinkrmalik]

Problems

1. Transcription cut off at the end (audio buffer truncation)

When using push-to-talk mode, releasing the hotkey after a brief pause would result in truncated transcription — the last few seconds of speech were lost.

2. ~1 second delay before mic activates

Pressing the hotkey had a noticeable lag before the microphone actually started capturing audio, making the app feel sluggish.

---

Root Causes

Audio buffer truncation

In AudioEngine.processAudioBuffer, audio frames were being discarded when silence detection or max-duration conditions fired. Both checks performed an early return before appending the current audio chunk to the buffer:

// BUG: max duration check returns BEFORE appending audio
if elapsed >= maxDuration {
    onMaxDurationReached?()
    return  // ← audio frame discarded
}

// BUG: silence detection returns BEFORE appending audio
if now.timeIntervalSince(lastSoundTime) >= silenceDuration {
    onSilenceDetected?()
    return  // ← audio frame discarded
}

// Audio only appended if we reach here
audioBuffer.append(...)

This meant:

1. The audio frame that triggered the condition was always lost
2. In push-to-talk mode, once silence duration elapsed, all subsequent audio frames with energy below the threshold were discarded (even though recording continued until key release)
3. The silence/max-duration callbacks fired repeatedly on every audio callback, not just once

Recording start delay

AppState.startRecording() called await soundManager.playStartSoundAsync() which blocked for the entire sound duration (up to a 1-second timeout) before activating the microphone:

// BUG: blocks recording start for ~1 second
if soundEffectsEnabled {
    await soundManager.playStartSoundAsync()  // ← waits for sound to finish
}
audioEngine.startRecording(...)  // ← mic only starts after sound completes

---

Fixes

AudioEngine.processAudioBuffer (AudioEngine.swift)

1. Reorder operations: Always append audio samples to the buffer before checking silence/max-duration conditions — no frames are ever discarded
2. Add silenceCallbackFired flag: Ensures the silence detection callback fires only once per silence period (resets when speech resumes)
3. Add maxDurationCallbackFired flag: Ensures the max duration callback fires only once per recording session
4. Both flags are reset in startRecording() for clean state on each new session

AppState.startRecording (AppState.swift)

• Start the microphone immediately — no more blocking on sound playback
• Play the start sound fire-and-forget after the mic is active
• Any brief sound bleed into the mic buffer is negligible and handled well by WhisperKit's noise model

---

Tests Added

6 new tests in AudioEngineTests:

• testStopRecordingWithoutStartReturnsEmpty — baseline behavior
• testSilenceCallbackFiresOnlyOnce — verifies the new dedup guard
• testMaxDurationCallbackFiresOnlyOnce — verifies the new dedup guard
• testAudioBufferNotEmptyAfterRecording — baseline recording works
• testAudioBufferPreservedWhenSilenceDetected — key test: buffer has data even after silence fires
• testAudioBufferPreservedWhenMaxDurationReached — buffer has data even after max duration fires

Testing

All 57 tests pass (swift test), including the 6 new AudioEngine tests. Manually verified both fixes work correctly — recording starts instantly and full audio is captured through key release.