[kixelated]

Problem

The scheduled macOS job cannot compile an_empty_gap_buffer_marks_the_next_samples: it spawns a non-Send capture producer. November 5's job confirms the failure before any tests execute.

Approach

Drive capture and its two reader assertions concurrently in the same task with tokio::select!. Completion of the assertions drops the producer; unexpected producer termination fails the test. Producer::run already documents its macOS local-task requirement, and the CLI runs capture in a LocalSet. No Send guarantee is added.

Keep the quest, updated with what remains and the websocket investigation. moq-dev/moq#7829 already reproduced the November 2 timeout on native macOS and fixed the client being dropped before server admission completed, with 100/100 native repetitions passing. Later scheduled runs fail before reaching it. moq-dev/moq#7667 owns independent failure reporting, not this compile fix.

Impact

• Public API: none.
• Wire: none.

Alternatives

An unsafe Send implementation would contradict the existing local-task contract. A LocalSet in this test is unnecessary: the producer and assertions can be polled together directly.

Choices made without asking

Retain the existing Linux PR gate. The existing native Platform check uses cargo check --all-targets and catches this error; it runs nightly and for release PRs, rather than on every main PR. Adding a macOS runner to every PR is not a cheap gap closure. A synthetic non-Send test marker would catch test spawns but leave production callers unchecked and remove the non-macOS Send assertion from test builds. Preserve the real platform types and contract checks. No workflow changes overlap with the reporting PR.

Validation

• nix run .#quest -- gate nix develop --command just ci origin/main: passed (exit 0), including all check phases, GitHub tooling tests, and 1,642 Rust tests (61 skipped).
• nix run .#quest -- check --base origin/main --branch quest/m0/nightly/macos: passed, 579 documents.
• git diff --check: passed.
• Native macOS validation remains scheduled after landing; the quest retains the week of scheduled-main websocket evidence wait.

Review

Claude Code / Claude Opus 5.5, medium effort, via quest review --base origin/main, session 065b8d07-0b42-4d8c-b4fe-209cfb22dcfa.

Round 1 requested consideration of a synthetic non-Send Linux test backend. Fixed the missing alternatives discussion in the quest, recording why that changes test-platform contracts without covering production callers. Round 2 reviewed ac64fc04a33bf69a93479117d50340057a5da60e: no findings. Head 7f0433be12afc2e04bc7deef9fd94fb689df9433 adds only a conflict-free merge with main; the capture and quest files are identical to that reviewed head. No new round is required under the merge exemption.

Landing decision

The maintainer authorized landing through the 2026-11-06 quest-complete session. The coordinator queues this PR. The scheduled validation wait remains in the quest rather than blocking this compile fix.

Follow-ups

The earlier direnv-hook test failure did not recur in this full local CI run. Consider isolating benchmark build-regression fixtures from the shared performance benchmark lock: the tooling phase spent hours queued behind unrelated benchmarks. No follow-up quest was created.

(Written by GPT-6.1 Sol with medium effort)

[kixelated]

Automated review by review (OpenAI)

Reviewed head: 7f0433be12afc2e04bc7deef9fd94fb689df9433.

Findings: No actionable correctness issues found in the capture-test fix or its quest update.

Fix status and direction:

• rs/moq-audio/src/capture.rs:859–865 removes the tokio::spawn Send bound by polling capture and both assertions in the test's own task. This matches Producer::run's macOS contract and the CLI's existing local-task handling without adding an unsafe Send guarantee.
• The empty gap marker still precedes two retained sample buffers. Both reader assertions remain required for success; premature producer completion panics, and successful assertions drop the capture future. The existing non-macOS Send assertions remain intact.
• Keep the scoped fix and the outstanding native-validation follow-up. A green scheduled macOS run and the subsequent week of websocket evidence remain unverified; the quest correctly stays open for them. No additional code change is requested by this review.

Scope and limits: GitHub-only static review of the two-file branch delta against live main at 1a6ac473d75fcbfd5d23bdbdfd64eb4750b077ad, including capture/fanout lifetime behavior, the ScreenCaptureKit handle, CLI local-task execution, and platform workflow wiring. The PR API also exposes a much larger file list against an older base; unrelated changes already present on current main were not re-reviewed. I did not run builds/tests or native macOS validation. The exact-head Check run was still in progress when inspected; the PR's local-test results are author-reported, not independently rerun here.

Rechecked immediately before submission: open, non-draft, unchanged head, and no existing published review to duplicate. This is a COMMENT review, not merge approval.

[pullfrog]

✅ No new issues found.

Reviewed changes

• Capture gap test driven on the test's own task — an_empty_gap_buffer_marks_the_next_samples replaces tokio::spawn(producer.run()) with a tokio::select! between producer.run() and the two reader assertions, panicking if the producer stops early. The reader branch completing drops the run future, which Stream::read documents as cancel-safe, so the device is released exactly as the old run.abort() did. The pattern matches production (rs/moq-cli/src/publish.rs drives capture.run() in the same select on the local task) and needs no Send or unsafe.
• Quest update — quest/m0/nightly/macos.md records the fix, keeps the quest open for the scheduled macOS evidence, and marks independent Nightly reporting as #7667's work. Its ## Waiting item names a check a later run can evaluate, and the cited #7829 fix is present in the tree (the websocket test was repaired, not skipped).

The capture-and-quest delta against current main is exactly these two files; the 1266-file PR list comes from the recorded base (40796ae) being on an older main lineage. I verified the 2026-11-05 job log: both reported errors come from the single capture.rs:859 spawn site, so this fix closes the compile failure rather than advancing past it.

  ^{｜ View workflow run ｜ Using deepseek-v4.1-flash-1017 (free via Pullfrog for OSS) ｜ 𝕏}