Phase 1b: guest audio out

Parent: #1. Depends on #2 (QemuSession). Can run in parallel with #5.

Requirements

• REQ-AUD-001: guest audio output is streamed to Moonlight

Out of scope: client microphone → guest. Sunshine and Moonlight have no client→host mic path (Sunshine’s mic_t is host audio capture). REQ-AUD-002 is deprecated. If Moonlight or Sunshine add mic upstream later, open a new issue.

Background (verified)

• QEMU: -audiodev dbus,id=snd0 plus an audio device (intel-hda + hda-output,audiodev=snd0, or virtio-sound-pci,audiodev=snd0), and -display dbus,audiodev=snd0. Our client calls org.qemu.Display1.Audio.RegisterOutListener(h) on /org/qemu/Display1/Audio with a socketpair fd (same pattern as the console listener) and exports org.qemu.Display1.AudioOutListener at /org/qemu/Display1/AudioOutListener.
• Listener methods: Init(id, bits, is_signed, is_float, freq, nchannels, bytes_per_frame, bytes_per_second, be), Write(id, ay data), SetEnabled(id, b), SetVolume(id, mute, ay per-channel 0-255), Fini(id). There can be several streams (ids). The NSamples property is the frame size (default 480, 10 ms at 48 kHz).
• Sunshine: src/audio.cpp:320 calls platf::audio_control(), then control->microphone(mapping, channels, sample_rate, frame_size, continuous, host_audio) (common.h:840), which returns a mic_t whose sample(std::vector<float>&) must block until a full frame of interleaved float PCM is ready. The Linux impl is PulseAudio in src/platform/linux/audio.cpp.

Tasks

• Routing: when capture == qemu, platf::audio_control() returns qemu::audio_control_t instead of the PulseAudio one. set_sink/sink_info become no-ops that return a synthetic sink so the rest of audio.cpp works unchanged. Keep the change to audio.cpp minimal; ideally none.
• qemu::mic_t: a ring buffer fed from Write on the GMainContext thread and drained by sample().
  • Convert any PCM format from Init (s16/s32/f32, LE/BE) to float.
  • Resample to Sunshine’s requested sample_rate if it differs (QEMU is usually 44.1 or 48 kHz; use FFmpeg swresample, already linked).
  • Remix channels to the requested mapping (stereo, 5.1, 7.1).
  • Mix multiple stream ids if present.
  • Apply SetVolume/mute.
  • On underrun, return silence when continuous is set, otherwise wait (match the PulseAudio impl’s semantics).
• Bounded buffer with drop-oldest on overflow, so a slow client can’t make latency grow without limit. Log overflow at debug level.
• Clean up on Fini and when the VM disconnects.

Tests

• Unit: fake QEMU sends Init (s16le 44100 Hz stereo) and Write with a known sine; assert sample() returns float 48 kHz frames with the expected frequency, RMS and channel layout. Also cover volume/mute, multiple streams, underrun, and overflow.
• E2E: guest plays a known tone (e.g. speaker-test -t sine -f 1000 in the Linux test guest). The test client decodes Opus and asserts the dominant frequency is 1000 Hz ± tolerance.

Done when

REQ-AUD-001 is linked to passing unit and E2E tests, audio works through a real Moonlight client (manual check, noted in the PR), and anvil requirement status passes. PR Closes #4.