[closetcaiman]

Closes #

⚠️ Breaking changes ⚠️

• N/A

Introduced changes

Bugfixes:

• fixed the iOS engine restart loop with iosVoiceProcessing: true. With voice processing (VPIO) on, starting a recorder made AudioEngine rebuild itself indefinitely.
  • the loop was initiated by spaghetti handling of RouteConfigurationChange posted by engine when initiating VPIO (AudioEngine.prepare)
  • the handler rebuilt the engine which again posted the notification and initialized VPIO -> the loop
    • regression from fix: lazy init of audio engine #1251 - fixing the dead go-to label revealed this
• fixed VPIO rendering errors:
  • with recorder-only graph, the VPIO unit (which is full-duplex) tried pulling from the graph which was never explicitly connected, so crashed with render err: -1
  • with AudioContext the mainMixer was created, but format was not adjusted to 48kHz required by the VPIO which violated the assumption of setVoiceProcessingEnabled and the engine stopped shortly after starting
    • from AVAudioIONode.setVoiceProcessingEnabled: "The output format of the input node and the input format of the output node have to be the same and they can only be changed when the engine is in a stopped state."

Changes:

• Route changes no longer restart the engine. Format changes restart the engine via AVAudioEngineConfigurationChangeNotification. A stopped engine with a tracked graph is marked for rebuild on its next start.
• VPIO output is connceted explicitly - connectVoiceProcessingIOOutput
• A failed start settles the state (settleStateAfterFailedStart: Paused / Idle, Interrupted kept for the interruption-end and foreground-retry paths) instead of leaving Running on a stopped engine.
• Recorder callbacks run outside the engine lock. Downgraded std::recursive_mutex to std::mutex.
  • the callbacks are returned as InputNotficationBlock - a workaround until the proper refactor, which then are invoked by the engine. Moving callbacks out of locks makes a possible ABBA deadlock impossible, so the previously introduced std::recursive_mutex is not needed and was a smelly workaround hiding the bigger bug.
• the conflicting options are settled silently - a proper refactor with explicit API throws/logging to come in a separate PR

Known issues:

• Siri + allowBluetoothHFP - iOS never sends our session an interruption end, and the foreground retry (fix(ios): recording resume after an interruption #1257) only runs after one, so recording stays stopped until the app restarts it.

Checklist

• Linked relevant issue
• Updated relevant documentation
• Added/Conducted relevant tests
• Performed self-review of the code
• Updated Web Audio API coverage
• Added support for web
• Updated old arch android spec file

[github-actions]

WPT non-regression comparison

PASS — no regressions · 0 improved section(s) · overall 3427 → 3427 (0)

Unchanged sections (28)

Spec section	Base pass	Head pass	Delta
Processing model	1	1	0
Other	56	56	0
AnalyserNode	138	138	0
AudioBuffer	140	140	0
AudioBufferSourceNode	220	220	0
AudioContext	58	58	0
AudioNode	261	261	0
AudioParam	680	680	0
BiquadFilterNode	275	275	0
ChannelMergerNode	30	30	0
ChannelSplitterNode	7	7	0
ConstantSourceNode	64	64	0
ConvolverNode	178	178	0
DelayNode	127	127	0
DestinationNode	0	0	0
DynamicsCompressorNode	4	4	0
GainNode	15	15	0
IIRFilterNode	87	87	0
MediaElementAudioSourceNode	0	0	0
MediaStreamAudioDestinationNode	1	1	0
MediaStreamAudioSourceNode	0	0	0
OfflineAudioContext	35	35	0
OscillatorNode	104	104	0
PannerNode	727	727	0
PeriodicWave	33	33	0
ScriptProcessorNode	0	0	0
StereoPannerNode	104	104	0
WaveShaperNode	82	82	0

_{Baseline: dba0ad8bb05caf741bb90c899857ee287b9378fb · Candidate: 050d477a91f0bc9a2c5305622ed92e58d41510fb}

_{Workflow run · this comment is updated on every push.}