Audio: RDPSND, AAC, and mute-on-minimize

How system audio reaches the client, the opt-in AAC compression path, and the minimize/refocus behavior.

System audio rides over the RDPSND virtual channel as 16-bit stereo PCM at 44.1 kHz. ScreenCaptureKit only supports 8 / 16 / 24 / 48 kHz, so the capture loop captures at 48 kHz and resamples to 44.1 with rubato before sending. The 44.1 kHz output is empirically load-bearing — the earlier 48 kHz feed over-fed mstsc into multi-second backlogs, and 44.1 fixed it (the precise mechanism is a pacing/rate-accounting effect; Windows resamples shared-mode streams to the endpoint mix format either way). A generation counter on the audio factory keeps a client reconnect from leaving a second capture loop feeding the channel. The vendored ironrdp-server carries a single patch that makes dispatch_server_events keep the newest queued waves on per-batch overflow instead of the oldest — without it, a one-off video-encode stall would bake a permanent audio-latency offset into the session.

The capture loop also self-heals a dead SCK audio stream: over a long session ScreenCaptureKit can stop delivering samples or transiently fail to start, which previously left the connection silent for the rest of the session (video is a separate stream and kept running). The loop now rebuilds the audio SCStream with capped exponential backoff (250 ms → 5 s) on both start failures and mid-stream end, resetting the backoff once a sample arrives; the generation guard still retires it on reconnect, so there's no double-capture.

AAC compression (opt-in, --enable-aac). By default audio is uncompressed PCM (~1.4 Mbit/s). Pass --enable-aac to encode it as AAC-LC over RDPSND (WAVE_FORMAT_AAC_MS, ~128 kbps by default — about 11x smaller), which matters over WAN or constrained links. The encoder is AudioToolbox (software AAC-LC); the wire payload is raw AAC access units. The server advertises AAC ahead of PCM, so clients that decode it (mstsc, Microsoft Remote Desktop / Windows App, FreeRDP built with AAC support) negotiate AAC automatically while clients without it fall back to PCM transparently. It's off by default because AAC adds ~40–50 ms of encoder priming latency — on a LAN, PCM's zero added latency is the better default. Tune the bitrate with --aac-bitrate (default 128000; 96000 saves the most bandwidth, 192000 is near-transparent for music).

Mute on minimize (default-on, opt out with --no-mute-on-minimize). When the client minimizes its window it sends the standard SuppressOutput { None } PDU; the server stops emitting both EGFX video frames and RDPSND waves until the client refocuses (RefreshRectangle / SuppressOutput { Some(rect) }). Without this, mstsc accumulates a backlog of video frames + audio waves during a long minimize that has to chew through on refocus, producing several seconds of input lockout, audio drift, and a video catch-up storm. With it, you get a brief audio gap on refocus and audio + video resume in sync. Both gates are debounced (1 s) so transient SuppressOutput flaps mstsc emits under wire pressure (e.g., during a heavy local cargo build) don't oscillate the mute and cause stutter. Pass --no-mute-on-minimize if you specifically want audio to keep playing while the client window is minimized — accepting that audio will drift by however long was spent minimized.