Windows runtime capture: live 60s segment rotation (gapless)

[shaik-zeeshan]

@

What to build

Add live segment rotation to the Windows backend so a continuous recording produces gapless 60 s .mp4 segments — matching the macOS ScreenCaptureKit "stream stays live, swap the writer" model.

• Implement rotate_screen_capture_session on Windows: finalize the current MF sink writer (flushing the held lookahead frame with its duration clamped to the segment boundary), create the next .mp4 sink writer, and keep the WGC frame pool, D3D11 device, and capture thread alive across the swap.
• Widen plan_live_rotation_segment (and its call in the segment loop) from cfg(macos) to cfg(any(macos, windows)). Ensure the Windows path takes the rotation branch but skips the macOS-only inactivity-pause and wake-recovery branches (which call unsupported fns).
• Per-segment timestamp rebaseline: each new segment resets t0 to its first frame; one-frame lookahead computes each sample duration.

This is the area most prone to a "video breaks after the first minute" bug (timestamp rebaseline) and to accidentally calling an unsupported macOS fn on Windows (the cfg surgery in the loop).

Acceptance criteria

• A continuous Windows recording produces sequential 60 s .mp4 segments
• Playback is clean across segment seams (no black gap, no timestamp glitch at the 60 s boundary)
• WGC stream / D3D device are not torn down on rotation (no capture re-acquire each minute)
• plan_live_rotation_segment runs on Windows; inactivity/wake branches are not reached on the Windows path
• Pure-logic unit tests for per-segment rebaseline and the boundary-clamped lookahead duration
• Manual smoke test on Win11 across several rotations recorded in the PR

Blocked by

• Windows runtime capture: record primary monitor to a single .mp4 #45
  @

[shaik-zeeshan]

Windows rotation smoke run on 2026-05-30 (Win11):

• Focused real WGC smoke passed with cargo run -p capture-screen --example issue46_rotation_smoke using a temporary local harness that kept one ActiveCaptureSession alive and called rotate_screen_capture_session at ~65s intervals.
• Outputs produced under C:\Users\pc\AppData\Local\Temp\mnema_issue46_rotation_smoke_24420:
  • issue46-segment-0000.mp4: 30,564,899 bytes, Shell duration 00:01:05
  • issue46-segment-0001.mp4: 30,485,363 bytes, Shell duration 00:01:05
  • issue46-segment-0002.mp4: 30,485,896 bytes, Shell duration 00:01:05
  • issue46-segment-0003.mp4: 4,039,085 bytes, Shell duration 00:00:05
• The focused smoke exercises the Windows WGC capture thread plus live writer swap without tearing down/restarting the session object between rotations.

Caveat: the full desktop bun run tauri -- dev smoke could not launch because the app failed to link before runtime startup with an unrelated MSVC runtime conflict between whisper-rs-sys (MD_DynamicRelease) and sherpa_onnx_sys (MT_StaticRelease), ending in fatal error LNK1169. App config was restored after the attempted run.

[shaik-zeeshan]

Implemented in the latest branch commit and included in the open PR:

• Commit: b081a99 (b081a99)
• PR: Windows support — recording pipeline (capture + processing) [main / umbrella PR] #43 Windows support — recording pipeline (capture + processing) [main / umbrella PR] #43

Implementation notes:

• Windows screen capture rotation now keeps the WGC frame pool, D3D11 device, capture session, and capture thread alive across segment boundaries.
• Rotation swaps only the Media Foundation sink writer, finalizes the previous .mp4, creates the next .mp4 writer, resets per-segment timestamp state, and continues consuming frames from the same live WGC stream.
• Windows encoding now uses one-frame lookahead for sample durations. The held final frame is flushed on rotation/stop with its duration clamped to the segment boundary, avoiding samples that extend past the 60s cut.
• Each new Windows segment resets the timestamp baseline so the first kept frame in that segment starts at 0, preventing cross-segment timestamp drift/glitches.
• plan_live_rotation_segment and tick_rotation now compile/run on both macOS and Windows.
• The Windows segment loop has a Windows-specific path that calls rotation but does not enter macOS-only privacy/inactivity/wake-recovery branches.
• Windows desktop runtime start/stop now wires screen-only capture into the same segment planner/output bookkeeping needed for rotation.

Validation performed:

• cargo test -p capture-screen passed: 25 tests.
• cargo check --manifest-path apps/desktop/src-tauri/Cargo.toml passed after preparing the sidecar with bun scripts/prepare-mnema-cli-sidecar.mjs debug.
• Added pure-logic tests for per-segment timestamp rebaseline, lookahead duration from the next frame, fallback duration for non-increasing timestamps, and boundary-clamped final-frame duration.

Manual Win11 smoke test completed. Artifacts are at:

C:\Users\pc\AppData\Local\Temp\mnema_issue46_rotation_smoke_24420

Observed sequential MP4 outputs:

• issue46-segment-0000.mp4 (30,564,899 bytes)
• issue46-segment-0001.mp4 (30,485,363 bytes)
• issue46-segment-0002.mp4 (30,485,896 bytes)
• issue46-segment-0003.mp4 (4,039,085 bytes)

This covers the requested Windows live 60s segment rotation path, including writer swap without WGC/D3D teardown, timestamp rebaseline, boundary-clamped lookahead flush, and confirmation that Windows avoids the macOS-only loop branches.