[Skeyelab]

Closes #544.

Real-time path (ADR 0002): human review of the diff. This changes Capture.receive, which runs on the audio thread. The change is atomics only: no locks, no allocation, no blocking calls (checked by a Code Reviewer).

The bug

receive reads joinRequests, then a few lines later ORs them into capturing. A stopFeeding (the Drive failed or retired) that landed between the two left the Copy fed with no writer draining it. After about 4 s its ring overflowed and the Take reported dropped frames that never happened. #872 closed the rest of this; this window was left.

The fix

• New stoppedFeeds mask, one bit per Copy. stopFeeding sets it first, before clearing the request and the capturing bit.
• Every block masks capturing with ~stoppedFeeds, so a stopped Copy gets at most the block in flight.
• The join block takes any Copy stopped since its request was read back out of capturing, so no stale bit is left for a later rejoin.
• attachMidTake clears the Copy's stopped bit; startCapturing clears all of them.
• A test-only hook, afterJoinRead (nil in the app), runs between reading the requests and acting on them. In the app it costs one nil check, only on a block with a join waiting.

Tests

• stopFeedingDuringJoinRead: the Copy is stopped from inside the hook. Red first: before the fix, the stopped Copy got 960 frames; now 0.
• rejoinAfterStopDuringJoinRead: a Copy that lost that race and joins again is fed only from its new join block, at its new join frame. Passed on its first run (it guards the stopped bit being cleared on rejoin).
• Not covered by a test: a rejoin landing between receive's two mask loads. The join-block clear above closes it; hitting it deterministically would need a second hook.
• Full ./scripts/verify.sh passed (spelling, vocabulary, package tests, macOS and iOS app builds).

Device Take needed (before External testers)

Checked by tests only. Ship in an [internal] tag (v1.1.16). Eric's Take from that TestFlight build: a Drive that fails as it joins mid-Take, with no false dropped frames in the report.

Follow-up noted, not fixed

Older than this change: joinRequests.bitwiseAnd(~requested) can wipe a new attach request that lands in the same block window, which needs a stop plus a reattach inside one block.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!