The quality programme (portfolio/quality/README.md) keeps a queue of open defects: every finding in a subject's latest audit that has no fix pull request recorded against it. A fix-mode session works from the top of that queue. A queue that still lists fixed defects is a problem: a session either fixes something twice or spends its time proving it is already done. A run on 2026-11-11 reported exactly that: PR #14327 had fixed world-clock-widget's high defect, nobody recorded it, and the queue still listed it.
What changed
python3 platform/ci/quality_audit.py queue now checks each row's subject in git. When commits have touched that subject's directory since the audit date, it says how many and shows the latest under the row:
[high] ventures/photo-editing (2026-11-09-photo-editing-r3.json#0, not-built-or-run): Public page says the renderers are dry-run placeholders ...
changed since audit: 7 commit(s), latest 3250444c7 photo-editing: optional numpy engine for the DNG developer ...
Cloud clones are shallow, and that needs two guards. The boundary commit of a shallow clone looks as if it touched every path, so it is dropped. And when the fetched history starts after the audit date, the row says "unknown" and prints the git fetch --shallow-since=<date> that would settle it, instead of reporting "untouched". --no-git skips the check.
What the first run found
With history back to 2026-11-04 fetched, 5 of the 24 open high-severity rows sat under a subject changed since its audit. One of them was a fix nobody recorded. The photo-editing audit of 2026-11-09 said the public page called the renderers dry-run placeholders and told users to run a shootflow command that is not installed. The page now says there is no installed command, gives the python3 -m shootflow.cli form, and describes the ffmpeg adapters as tested only on synthetic media. That fix came in with audit round 30 (PR #14329), whose commit title mentions it. The audit file now lists that PR in fix_prs, and the queue is down to 23.
A first draft of the check hid commits titled "quality: audit round ...", on the theory that audit commits only add audit files. That draft would have missed this exact fix. An audit commit shows up for a subject only when it changed the subject's own files, so it now counts like any other.
The other four flagged rows are ordinary product work (new photo-editing features, open-tennis-robots sizing, quick-pick-survey measurements) and a red-team usage-tracking change on caa-record-check. None of them obviously fixes its queued defect, so they stay open. A commit under a subject means "look first", not "fixed".
Why software and not a rule
The README already told fix sessions to record their PR in fix_prs, and sessions mostly do. The misses come from fixes that land in some other run: a polish pass, an audit round that fixes what it finds, a product session. Those runs have no reason to open an old audit file. A line of output at the moment someone reads the queue catches that case better than another sentence in a guide.
Tests: platform/ci/test_quality_metrics.py, ChangesSinceTest (a fake git covers the boundary commit, a history that starts too late, and no git).