ci(verify): export real-capture Studio recordings and check them against their tracks; fix audio late in exports of never-opened camera projects

[richiemcilroy]

Behavior

The real-capture suite now exports each non-crash Studio recording and checks the export against that recording's source tracks. It verifies that required screen pictures remain present, appear on the decoder's expected frame grid, and that microphone markers remain present and within 5 ms of their recorded times.

This also fixes exports of camera recordings that have never been opened in the editor. Those projects can have no saved clip offsets. ExporterBase now fills them with the same initial configuration the editor uses, while preserving any offsets already saved. In the measured failing recording, all 59 microphone markers were 169.5 ms late before the fix and aligned after it.

Exact timing checks

The oracle independently models request rounding, source frame labels, and track offsets. When consecutive pictures share a frame label, the previous picture can occupy that cache slot; the new picture must appear at its first label or the immediately following request. That next request need not correspond to a source sample. Later source labels do not justify a later first appearance.

A marker-index lookup avoids repeatedly scanning all earlier pictures during long soaks. Grouping 54,000 decoded samples from 27,000 distinct pictures took 10.2 ms locally; twice that input took 16.0 ms.

Validation

• Replayed the exact failed macOS startup and Windows long-recording artifacts through both AVFoundation and forced FFmpeg on macOS. The old oracle reproduced all 3 and 22 reported timing failures respectively. The revised oracle reports zero timing failures and zero missing required pictures on those same exports.
• Restoring the old f32-floor decoder produces two timing failures and one missing picture in the saved macOS recording; the revised oracle catches them with both decoders. The product decoder was restored afterwards.
• Four focused regression tests cover label collisions with and without a source sample at the next request, rejection of later labels, exact non-colliding labels, and half-frame ties. Restoring the old oracle behavior makes the new tests fail.
• cargo fmt --all, cargo check -p cap-verify, strict Clippy, 12 recording unit tests, 10 verifier library tests, and 45 exporter library tests pass locally.
• Fresh local real capture on the final head passed studio-camera/startup-drops, studio-camera/long, and studio-screen/late-mic: zero missing required pictures in all three and microphone markers aligned to the recorded times.
• Final-head CI, full selected engine suites, A/V sync on all three operating systems, and native real capture on macOS and Windows passed. Greptile reviewed 006ec61e4 at 5/5 with all review threads resolved.
• The exact merge tree with the performance-cache fix passed the complete engine suite in run 39129950471.
• With local crates rebuilt independently for both sides, the basic-export repeat against current main measured CPU +0.0% and wall +1.5% over three interleaved pairs (run 39130619110). The earlier basic-export slowdown did not reproduce. The complete guarded comparison measured candidate-relative CPU changes of about +0.4% for jitter export, -1.2% for seeking, and +4.6% for playback (reverse comparison 39128980042; candidate is its base).
• The guarded comparison with 0.6.0 measured lower CPU for jitter export, seeking, and playback. Basic export was not comparable there because correctness outcomes differ. The existing Linux camera-crossfade mismatch and host-specific Instant timing baseline finding remain separate open issues; this PR does not claim whole-product regression freedom.

The marker-detector prerequisite, #2694, is merged.

Confidence Score: 5/5

The PR appears safe to merge; no new actionable issue was established.

Summary

This PR checks finished Studio exports against their recorded screen pictures and microphone markers. It also fixes audio timing when a camera project has never been opened in the editor.

• Studio exports fill in missing track offsets before rendering.
• Studio recordings now get checked against their exported video and audio.
• Audio marker checks ignore near-silent codec noise.

The previous threads were rechecked. Missing microphone markers inside the checked range are reported, request rounding follows the decoder’s rule, and picture lookups no longer scan every earlier picture. None of the supplied threads has a Comment number, so previousFindings is empty.

Review was based on code inspection. No builds or tests were run.

_{Reviews (7) · Last reviewed commit: "chore: merge marker detection fix for ex..." · Reviewed by Greptile}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
cap-status	Ready	Preview	Oct 18, 2026 12:21pm UTC

[capy-ai]

Capy found 1 potential issue (1 medium).

View 1 other finding in Capy.

[richiemcilroy]

@greptileai review

[capy-ai]

Capy found no issues.

[richiemcilroy]

@greptileai review

[capy-ai]

Capy found no issues.

[richiemcilroy]

@greptileai review

[capy-ai]

Capy found no issues.

View 1 other finding in Capy.

[richiemcilroy]

@greptileai review

[capy-ai]

Capy found no issues.

[richiemcilroy]

Blacksmith macOS 26 flagged 13 pictures in studio-camera/long. Reproduced with that run's recording on a MacBook Pro, with both decoders: each picture spans two frames 1/60 s apart on a 30 fps track, so a picture's first frame can share its grid label with the previous picture's second frame (2.983 s and 3.000 s both label 90). The decoder may answer that label with either, so the next picture first shows one frame later, at its own second label. That is the contract, so the check now allows any of a picture's labels when its first one is shared with the previous picture. Local cases still pass, and the f32-floor decoder mutant still fails (5 and 6 pictures). @greptileai review

[capy-ai]

Capy found no issues.

[richiemcilroy]

hey @greptileai, please re-review the PR

[capy-ai]

Capy found no issues.