What happens
scan-matrix.yml fails every week on main. The most recent, 2026-08-24:
conformance (32) failure
conformance (64) failure
conformance (128) failure
block=64: something in this sweep wants a human — see the summary
That is the workflow working as designed — a violated prediction is a
finding, and the exit code is how it asks for a human. Nothing is broken.
Why it is still a problem
The findings it is reporting are already filed:
All three have been open since 2026-08-22. So every weekly run re-reports the
same three known things by turning main red, and will keep doing so until the
corpus is re-recorded. A repository whose default branch is permanently red
teaches everyone — including its maintainer — to stop looking at the badge, and
the next new finding arrives into a signal nobody reads any more.
This is the same failure mode as
vyncint/termlens#159: a check
that is right about the facts and wrong about what to do with them.
What would fix it
A recorded baseline. Accepted, already-filed findings go in a file the sweep
reads; the run fails only on a finding that is not in it. Then red means "this
is new", which is the only thing red can usefully mean. Re-recording the corpus
(#11) would clear today's three, but the mechanism is what stops the next three
doing this again.
Alternatives, both worse: dropping the exit code loses the signal entirely, and
muting the schedule until the backlog clears just moves the problem to whenever
someone forgets to unmute it.
Done when
main is green while every known finding is filed, and a new finding still
turns it red.
What happens
scan-matrix.ymlfails every week onmain. The most recent, 2026-08-24:That is the workflow working as designed — a violated prediction is a
finding, and the exit code is how it asks for a human. Nothing is broken.
Why it is still a problem
The findings it is reporting are already filed:
corpus-drift)lane-env-barrier-below-is-harmlessno longer shows what it wasAll three have been open since 2026-08-22. So every weekly run re-reports the
same three known things by turning
mainred, and will keep doing so until thecorpus is re-recorded. A repository whose default branch is permanently red
teaches everyone — including its maintainer — to stop looking at the badge, and
the next new finding arrives into a signal nobody reads any more.
This is the same failure mode as
vyncint/termlens#159: a check
that is right about the facts and wrong about what to do with them.
What would fix it
A recorded baseline. Accepted, already-filed findings go in a file the sweep
reads; the run fails only on a finding that is not in it. Then red means "this
is new", which is the only thing red can usefully mean. Re-recording the corpus
(#11) would clear today's three, but the mechanism is what stops the next three
doing this again.
Alternatives, both worse: dropping the exit code loses the signal entirely, and
muting the schedule until the backlog clears just moves the problem to whenever
someone forgets to unmute it.
Done when
mainis green while every known finding is filed, and a new finding stillturns it red.