The tk .got capture files are invisible to testmgr's privatization

Found 2026-08-19 by Track T (plexus-T), on the first run of the lint added for [[bug-t-split-jobs-misses-a-tmp-path-reached-through-a-shell-variable]]. Filed for routing — T owns the tool, the owning lane owns the recipe.

What is still there

Makefile:376-381, inside test-nilpy:

for t in tkinter_facade:$(TESTTMP)/test_nilpy_tkinter26 \
         field_class_identity:$(TESTTMP)/test_nilpy_fldcls26 \
         callbacks:$(TESTTMP)/test_nilpy_tkcb26; do \
  src=$${t%%:*}; bin=$${t##*:}; \
  timeout 120 xvfb-run -a $$bin > $(TESTTMP)/$$src.got 2>&1 \
    || { echo "  tk: $$src EXITED NONZERO under Xvfb"; cat $(TESTTMP)/$$src.got; exit 1; }; \
  diff -u examples/tk/$$src.expected $(TESTTMP)/$$src.got \
    || { echo "  tk: $$src OUTPUT CHANGED"; exit 1; }; \
done; \

9f11b405d fixed the binaries: the item list now spells each one by full path, so union-find sees the literals and the three jobs merged into one ordered job. That was the right fix and it closed the callbacks red.

The .got capture files were not part of it. $(TESTTMP)/$$src.got becomes /tmp/$src.got in make -n, and testmgr's tmp_re (/tmp/[A-Za-z0-9_./+-]+) does not match it — $ is not in the class, so the token is not merely wrong, it is absent.

Why it still matters with the merge already fixed

The merge half is genuinely fixed; this is the privatization half, which is a different mechanism with a different consequence:

• testmgr rewrites literal /tmp paths in the recipe text it executes, so two concurrent runs get their own scratch. It cannot rewrite what it cannot see.
• So /tmp/tkinter_facade.got, /tmp/field_class_identity.got and /tmp/callbacks.got are shared across every concurrent run on the box — and this box runs two clones' testmgrs by design (the watcher's plus a dev checkout), which is the documented co-tenancy in devdocs/dev/track-t.md.
• The failure it produces is tk: <src> OUTPUT CHANGED — a diff against the expected file, i.e. it presents as a NilPy tk regression rather than as a collision. Same shape as the UDP port collision recorded in track-t.md: the loser of a race on a shared global does not fail as a hang, it answers with someone else's data and the red arrives with a plausible wrong diagnosis attached.

Not yet observed in the wild. Filed because it is cheap to close and expensive to triage, and because the class has now bitten this exact recipe once already.

The fix

Same shape as the one that landed: state the path in the recipe rather than building it from a variable. The item list already carries src:bin; a third field, or a second full-path spelling, makes the .got visible to the scan:

for t in tkinter_facade:$(TESTTMP)/test_nilpy_tkinter26:$(TESTTMP)/tkinter_facade.got \

Then got=$${t##*:} and use $$got. Any spelling works as long as every /tmp path the recipe touches appears literally in make -n output.

Guard

tools/testmgr_tmp_var_devtest.py carries this as its one KNOWN entry, so it prints on every run rather than being silently tolerated. When this is fixed, delete that entry — the devtest already prints an explicit "no longer reaches /tmp through a variable — remove it from KNOWN and close its ticket" when a known-open instance clears.