Process runs in this repository pick their next fix from platform/ci/friction_rank.py. It groups the complaints that runs record about instructions and tools, then hides the ones a later commit has probably fixed. To see those commits on a shallow cloud clone, it first fetches main back to the oldest report.
What went wrong
This run started promptctl.py measure in the background, as the process prompt says to, and ran the friction ranker next to it. Both fetch main into the same shallow clone. Git lets one fetch at a time hold .git/shallow.lock, and a second fetch that meets the lock fails at once. The ranker treated the fetch as best effort and kept going without mentioning it.
The output looked normal. It printed "fetching main since 2026-11-13", then marked every row "last report predates the fetched history: touched-since unknown". Five of the top fifteen rows were allocate.py complaints that later commits had already dealt with. A run record written earlier today reports the same thing ("--live printed 'last report predates the fetched history' for every cluster although it said it fetched since 2026-11-03"), and that run found its target by reading the records by hand.
Once measure had finished, the same command (--live --ungated --days 3) hid 48 clusters as touched since their last report, up from 26, and no allocate.py row was left in the top eight.
What changed
deepen() now waits up to 240 seconds for another fetch's shallow.lock to clear before it fetches. If a fetch still loses the race, it retries once. If the fetch fails for any other reason, it prints why to stderr, for example (fetch failed: fatal: unable to access ...; touched-since stays unknown for older reports, so --live hides less than it should). A new unit test drives the lock, the retry and both failure messages with a fake clock. All 24 tests in test_friction_rank.py pass.
The same run added MUJOCO_LOG.TXT to the root .gitignore. Quality census runs that exercise MuJoCo projects from the repository root were leaving that file behind, and the stop hook flagged it as untracked work.
What's left
promptctl.py measure also swallows a failed fetch. It lives under prompts/, which needs the owner's approval to change, so that fix is recorded in this run's notes rather than made here.