check-pr-size.py cannot classify .agents/scripts/**, so no PR can add a job script there

[localai-org-maint-bot]

Row: UPSTREAM-SYNC-HEADPIN

scripts/check-pr-size.py cannot classify .agents/scripts/**, so no pull
request may add or modify a file there.

Reproduction

Committing two lease-job scripts beside the three already tracked in
.agents/scripts/:

$ python3 scripts/check-pr-size.py --base origin/main --head HEAD --branch row/UPSTREAM-SYNC-HEADPIN
ERROR: PR size check could not classify the change: unclassified repository path '.agents/scripts/headpin-e126687-job-a.sh'
$ echo $?
1

classify_path fails closed at scripts/check-pr-size.py:519 and no rule above
it matches .agents/scripts/. Failing closed is right; the gap is that the class
was never taught.

Why it matters

Three job scripts are already tracked there, and
.agents/specs/glm5-next-flash.md says why: a run is committed "so the run is
reproducible". .agents/specs/cuda-keepquant-iq2xs-iq4xs.md already records this
same checker rejecting .agents/scripts/glm53-dsa-first-load.sh, so this is the
second time it has been hit.

The effect is that evidence a report depends on cannot be committed. #2593's
sync report §5.3 rests on what changed between two job scripts; its C10 now
records that the scripts exist only as sha256s and that the diff is therefore
checkable only against itself.

What this is not

Not a request to widen the classifier by a general directory exemption. AGENTS.md
requires policy, checker, document, script, test, CI, generated and product paths
to be classified explicitly and forbids hiding mutable files behind a directory
exemption. .agents/scripts/** wants its own explicit class.

Changing it is a semantic checker change, so it needs a spec, a red-before test
and green-after evidence, and it must not ride inside a sync record. That is why
#2593 reverted the attempt rather than fixing it in flow.

[localai-org-maint-bot]

Fixed on main at 6d1335568 — "classify .agents/scripts, whose absence left check-pr-size's own test red on main" (#2609, same defect).

.agents/scripts/<name>.{sh,py,hip,cu} now classifies as evidence, on the reasoning the checker already records for docs/bench-evidence's own .sh and .cu — the recipe that produced the number. Verified rather than assumed before choosing that class: nothing under CMakeLists.txt, .github/, scripts/, src/ or include/ references .agents/scripts/; the only referents are spec files.

Red-before: 1 failed, 52 passed, 157 subtests. Green-after at HEAD:

53 passed, 157 subtests passed in 0.27s

and check-pr-size.py --base origin/main --head HEAD returns OK on a change that adds a file there.

It is not a directory exemption, which the protocol forbids: the pattern names an extension set and a single path segment, so .agents/scripts/nested/evil.sh, .agents/other.sh and dotfiles all still fail to classify.

Closing as a duplicate of #2609. #2588 was already closed for the same defect — that made three issues for one bug, which is worth noting for whoever triages next.