The crtl census collides with itself in a shared scratch dir

test/crtl_declaration_census.sh takes its tmpdir from the caller (Makefile:35714 passes $(TESTTMP)) and writes three FIXED names into it:

SRC=$TMP/crtl_census.c
BIN=$TMP/crtl_census
LOG=$TMP/crtl_census.log

No mktemp, no per-job suffix, no lock. The verdict is then

if grep -q 'does not define' $LOG; then

so a second invocation sharing that $TMP gives one run's grep the other run's log, and $BIN is executed while the other run may be rewriting it — which produces "compiled clean, then no stdout", with no error anywhere.

Why this is filed WITHOUT a reproduction

It would be easy to attach this to the lib-test#55 red of 2026-09-17 and wrong. That red is fully explained by something else: the row runs $(PXX_STABLE), pin v410 predates 984be7e19, and the pinned compiler genuinely did not define c_pthread_create. It reproduces alone, in a fresh mktemp -d, with nothing else running — so contention had no part in it.

What is claimed here is only what the source shows: fixed names in a caller-supplied directory, with no exclusion. Whether the tier ever runs this script twice concurrently is not measured and is the first thing to check — if it does not, this is latent rather than live, and the fix is still cheap.

Positive control the fix needs

A guard here must be able to FAIL, and the honest control is two deliberately concurrent invocations against one $TMP, asserted to both pass. Run against the CURRENT script first: if that does not fail, the hazard is latent and the ticket should say so rather than being closed as fixed.

The fix

TMP=$(mktemp -d "$2/crtl_census.XXXXXX") — keep the caller's directory as the PARENT so the harness still controls placement and reaping, and make the three names unique per run. Do not add a lock; there is no reason these runs need to serialise, only to stop sharing names.