PlatformAudio stops working after the last PeerConnection is released: AdmProxy::Terminate tears down the platform ADM #1468

Summary

Once the last PeerConnection in the process is destroyed, PlatformAudio stops working: the next PlatformAudio::new() logs -1 recording devices, -1 playout devices, and recording_devices() never returns.

Cause

The media engine is reference-counted per PeerConnection (ConnectionContext::MediaEngineReference): WebRtcVoiceEngine::Init() runs when the first PeerConnection is created, and WebRtcVoiceEngine::Terminate() when the last one is destroyed, which calls adm()->Terminate().

AdmProxy::Init() is a no-op (the synthetic and platform ADMs are initialized in the constructor), but AdmProxy::Terminate() forwards to both sub-ADMs. After one Terminate/Init cycle the platform ADM stays uninitialized: RecordingDevices() / PlayoutDevices() return -1, and PlatformAudio::recording_device_count() casts that i16 to usize, so recording_devices() walks 2^64 - 1 indices at 100% CPU. The synthetic ADM loses its playout queue the same way.

Where the Terminate comes from (macOS, lldb, worker thread):

PeerConnection::~PeerConnection
  -> ConnectionContext::MediaEngineReference::~MediaEngineReference
  -> WebRtcVoiceEngine::Terminate
  -> AdmProxy::Terminate
  -> AudioDeviceModuleImpl::Terminate
  -> AudioDeviceMac::Terminate

When it is reachable

Whenever closed PeerConnections are actually released and no other PeerConnection is alive, for example:

1. PlatformAudio::new() and publish a track created from audio.rtc_source()
2. room.close() the only room and drop it
3. Use PlatformAudio again: recording_devices(), or connect a new room and publish the microphone -> hangs

We hit it in a fork that carries #1405, which lets the room's graph be released after close. With #1436 / #1435 it also follows server-ended rooms. It stays hidden while any PeerConnection is kept alive (a leaked one, or a second room connected in the same process).

What we did

• AdmProxy::Terminate() no longer terminates the sub-ADMs; they live as long as the proxy and are still terminated in ~AdmProxy. The voice engine already stops playout and recording and unregisters its callback before it calls Terminate(), so there is nothing left to stop there. Re-initializing them in Init() would also work, but would have to restore device selection and the playout/recording mode state.
• PlatformAudio treats a negative device count as zero instead of casting it.
• Regression test: connect, publish a PlatformAudio track, close the last room and wait until it is dropped; the recording-device count must be unchanged and the next room must publish the microphone again (before the fix: 5 devices, then 0).

Environment: macOS 26 on Apple Silicon, libwebrtc webrtc-51ef663 (m144), livekit-server 1.13.6 and LiveKit Cloud.