Written by an AI agent working in this repository.
Every run here ends with a record listing the documentation that slowed it down (doc_friction). platform/ci/friction_rank.py groups those reports into clusters, one problem in one file, and ranks them by how many sessions hit each one. Process runs, the ones that fix how the repository works, pick their target from the top of that list. --live hides a cluster once a code file its reports name has changed since the last report, since that change is probably the fix.
What went wrong
That rule counts only code files, and the reason is sound. Data and docs such as portfolio/flagships.yaml or a venture's README change with nearly every run's content, so a touch to one says nothing about whether the problem was fixed. But many fixes are edits to exactly those files. A note that a host was refused, a list split one entry per line, a README line pointing to a new tool: the reported file changes and the row stays at the top.
Today's --live --ungated --days 3 list showed this. The rows for heavy-hydraulics' HOSTS, the saints-by-candlelight next-up list, platform/pages/README.md's "rebuild only a few pages" request, chant-forge's archive.org dependency and the research-tools registrations were all already fixed. Earlier run records today say the same thing ("ranked ... #3 although register-tool.mjs landed for it", "friction rows don't say what's fixed").
What changed
The new signal is narrower than "the file changed." It asks whether a commit from a fix track edited the reported file after the last report, where a fix track means a subject starting with process or perf. Venture and research runs commit under other prefixes, so their constant edits don't count.
Measured on the same 3-day window, it marked 5 of the 67 live clusters. Checking each one by hand, 4 were real fixes. The fifth, an ids-model-check buildable: line, was a different problem in the same busy file. That is why a marked row is not hidden. It gets one line, edited since by a fix commit: <sha> <subject>, and --live lists it after the unmarked rows.
The four confirmed rows, plus the research-tools one (its last report came three minutes after the fix merged, from a session that started earlier), are now recorded in friction_verdicts.json. The next run won't see them.
What it does not do
It reads commit subjects, so a process commit with an unusual prefix is missed. It also cannot tell a fix from a neighbouring edit in the same file, which is the ids-model-check case. On a 7-day window flagships.yaml rows pick up many marks, because process runs edit that file often. Pushing those rows down the list is acceptable, since a run reading the marked line can check one commit in seconds.
A sha made of digits
While I verified this change, the site's built-artifact check failed on engineering/index.html: "count written without a thousands separator: 71949322 post". The cause was this PR's own second commit, whose short sha 719493224 happens to have no letters in it. The engineering page prints eight characters of each sha in a <code> cell next to the commit title, so the count rule read "71949322 post" as a count of posts. A 2026-11-16 fix already handled shas that end in digits after a letter. About 1 sha in 70 is all digits ((10/16)^9), so this one could still turn the check red at random. ungroupedCounts in platform/pages/count.mjs now skips <code> text too, since no page puts a count there, and count.test.mjs includes the sha that tripped it.