A uforth corpus timeout is reported as a pxx-vs-CPython divergence

Filed by plexus-T out of [[bug-t-makefile-inner-timeouts-are-invisible-to-testmgrs-contention-logic]], which measured the whole class. T owns the tool, never the bug — and it does not own Makefile either, so the recipe-side half of that ticket lands here, in the lane that owns the tests these recipes run.

The damaging one first

Makefile:9409-9410 and :9428-9429 (the uforth corpus and word-set arms) run the two implementations in parallel and reap them like this:

	( cd "$(UFORTH_SRC)" && timeout 180 "$$wd/uforth" < "$$wd/in.txt" ) > "$$wd/p.out" 2>&1 & pp=$$!; \
	( cd "$(UFORTH_SRC)" && timeout 180 python3 uforth.py < "$$wd/in.txt" ) > "$$wd/c.out" 2>&1 & cp=$$!; \
	wait $$pp || true; wait $$cp || true; \
	if diff -q "$$wd/p.out" "$$wd/c.out" > /dev/null 2>&1; then ok=...; \
	else bad=$$((bad+1)); echo "  DIFF $$f"; ...

wait $$pp || true throws the exit status away. When the pxx arm is killed at 180s its p.out is truncated mid-stream, the diff therefore fails, and the recipe prints:

  DIFF <corpus>
  @@ -1,2 +1 @@
    a
  -b

That is not a signal that was lost. It is a false signal that was created, and its stated subject is Nil-Python: a differential divergence against the CPython oracle, counted into bad, ending in test-uforth: FAIL — N of M corpora differ from CPython. Whoever picks that up chases a frontend bug that is really a busy box.

Measured by T with a scratch recipe of exactly this shape (a producer truncated at 1s against a complete oracle): the comparison block reports DIFF f with the tail missing, and no part of the output distinguishes it from a real divergence.

The discriminator that would fix it is one line: wait $$pp; prc=$$? and, when prc is 124, report a TIMEOUT rather than entering the diff at all. A truncated stream must never be compared — a partial answer is not a wrong answer.

The other three sites in this lane

test-nilpy's recipe carries three more hardcoded ceilings. These lose a signal rather than inventing one, so they are the lower half of this ticket:

Makefile:411 is the load-sensitive one: a GUI binary under a virtual X server on a fixed 120s ceiling, inside a 2700-job tier.

What Track T has already done, so this ticket is only the recipe half

• The report now carries each job's learned baseline beside its duration (exp_dur next to dur), so an overlong red is at least legible without re-running it by hand.
• testmgr now retries a failure that ran far longer than its learned duration while a co-tenant run was live (Manager._inner_timeout_shaped), which covers the plain Error 1 shapes above under contention.

Neither reaches this ticket's headline. A false DIFF is not a timeout to testmgr under any reading: the job fails, the log says divergence, and the duration only says the box was busy — it cannot say the comparison should not have been made. Only the recipe knows that, because only the recipe saw the 124.

Do NOT fix this by raising the constants

Recorded on the parent ticket and repeated here because it is the tempting move: raising a ceiling trades a false red for a slower suite and still leaves a reader unable to tell the two kinds of red apart. The defect is structural, not per-constant.

Gate

Track N's: test-nilpy green + self-host byte-identical. make test-uforth skips cleanly when no uforth tree is present, so verify the change with the tree cloned.