Windows microphone capture to openable .m4a segments (audio capture seam)

[shaik-zeeshan]

What to build

The foundational tracer bullet for Windows audio capture: record the microphone end-to-end into openable, rotating .m4a segments, decoupled from screen capture.

This slice establishes the shared audio seam that the system-audio and inactivity slices build on:

• Introduce a platform-neutral AudioCaptureSession trait (rotate→finalization, stop, is_live, take_stop_error), mirroring ScreenCaptureSession. The Recording Lifecycle holds active_microphone_session and active_system_audio_session as Box<dyn AudioCaptureSession>. macOS microphone keeps its concrete type for now; the trait's finalization type must reconcile with the existing macOS finalization consumed by microphone output finalization.
• WASAPI default-device microphone input → AAC .m4a via Media Foundation (AAC encoder + MPEG-4/M4A sink in the writers crate). Mirrors the macOS separate-audio layout — zero schema churn.
• Reuse the existing SegmentPlanner layout verbatim: <save_root>/YYYY/MM/DD/audio/<session_id>-segment-####.m4a, microphone via a mic_session-prefixed id. Populate SourceSessions.microphone source-session metadata (currently never populated on Windows).
• Segment rotation on the 5-minute Capture Segment Duration boundary, driven by the existing source-independent wall-clock CaptureClock + SegmentSchedule — no new timing source. Add the Windows microphone rotate arm to the rotation tick.
• Audio-only restructure: the rotation tick must not hard-require a screen planner/session. The "no screen planner ⇒ fail session" guard and the screen-capture-stopped failure check become conditional on whether screen is an active source (a no-op for macOS, which always has a screen planner when audio is active). No per-segment screen workspace dir is created for audio-only sessions.
• Finalization validation: define "openable .m4a" as "Media Foundation Source Reader opens it and MF_PD_DURATION > 0", reusing the existing injectable has_positive_duration validator seam (un-gate the structural helpers to also build on Windows; only the MF probe leaf is new). Matches the macOS audio contract.
• Commit produced Audio Segments, but do not enqueue audio processing jobs on Windows yet (capture-and-store now; a future audio-decode slice will backfill them). Non-goal: audio decode-to-PCM.
• Flip supported_sources.microphone on for Windows so the source becomes user-selectable.

The Windows user pause/resume control stays hidden until the inactivity slice fixes resume (see slice 3c).

Non-goals (explicitly out): audio decode-to-PCM for transcription/speaker/speech-activity, frame-index sidecars, exact/scrub previews, microphone device-change handling (slice 1b), permission UX (slice 1c), VAD/activity emission (slice 1d).

Acceptance criteria

• AudioCaptureSession trait exists; runtime holds microphone (and system-audio) sessions as Box<dyn AudioCaptureSession>; macOS finalization reconciled
• Microphone-only recording on Windows produces openable .m4a segments under the shared audio/ dir with mic_session-prefixed ids
• SourceSessions.microphone metadata is populated on Windows
• Segments rotate on the 5-minute boundary using the existing CaptureClock/SegmentSchedule; no screen session is required
• Audio-only sessions no longer fail for lack of a screen planner; macOS behavior is unchanged
• .m4a finalization validated via MF Source Reader positive-duration probe through the existing injectable validator seam
• Audio Segments are committed but no audio processing jobs are enqueued on Windows
• supported_sources.microphone is true on Windows and reveals the mic source in the UI
• Windows user pause/resume control is hidden (until slice 3c)

Blocked by

• Spike: verify wasapi 0.23 capture-clock timestamps + peak-since-last-poll level reads #51

[shaik-zeeshan]

Implemented and smoke-tested on Windows. ✅

PR: #43 (branch t3code/windows-support-gating)

Commits:

• 40cd04b — implementation: AudioCaptureSession seam, WASAPI→MF AAC/M4A backend, audio-only rotation tick, capability/UI gating
• 6213e8c — on-device smoke-test example (crates/capture-microphone/examples/smoke_mic_m4a.rs)

Smoke test result

Ran the real WASAPI default-endpoint capture + Media Foundation AAC/M4A sink end-to-end (record → rotate a segment → stop → validate):

Check	segment 0	segment 1
mic_session-prefixed .m4a created	✅	✅
Size	50,051 B	50,161 B
MF Source Reader MF_PD_DURATION > 0	✅	✅

Cross-checked the output with ffprobe (independent demuxer, not our own probe):

codec_name=aac   sample_rate=48000   channels=2
format_name=mov,mp4,m4a,3gp,3g2,mj2
duration=3.029313
header: ....ftypmp42....mp41isom...

Confirms genuine, openable AAC-in-M4A — stereo 48 kHz, duration matching the ~3 s capture. microphone_permission_state reports Granted; is_live toggles correctly across start/stop; finalization records carry the expected output_file with no spurious discard.

All 9 acceptance criteria met; compile-verified on Windows (cargo check --all-targets, bun run check clean) with macOS paths kept behavior-identical.