gap-suite (2): test_gap_9676_stdin_unref_ref_keeps_reader cannot fit the suite's uniform 10s timeout (Node oracle itself times out) #9783

gap-suite (2) is red on main, and the regression it reports is not a Perry defect —
it is the Node.js oracle timing out on a fixture whose own minimum runtime exceeds the
suite's per-test cap. Filing so the fixture's author can pick a fix; I have deliberately not
changed it.

What CI reports

SKIP  test_gap_9676_stdin_unref_ref_keeps_reader (Node.js TIMEOUT (killed after 10s))
[82/233] (00:12:51) test_gap_9676_stdin_unref_ref_keeps_reader … NODE_FAIL

REGRESSIONS — these were expected to pass:
  - test_gap_9676_stdin_unref_ref_keeps_reader: pass -> node_fail

Why it cannot pass

test-files/test_gap_9676_stdin_unref_ref_keeps_reader.ts drives four child roles
sequentially:

const roles = ["unref-ref", "unref-ref-churn", "unref-only", "pause-resume"];
...
for (const r of roles) await run(r);

and each role's run holds open a hardcoded timer before it finishes:

// Late enough that the churn role has finished collecting first.
setTimeout(() => {
  ...
}, 2500);

That is 4 x 2500 ms = 10 s of intentional delay alone, before any process-spawn
overhead — against run_parity_tests.sh:58:

PERRY_RUN_TIMEOUT="${PERRY_RUN_TIMEOUT:-10}"

a single uniform 10-second wall-clock budget applied to every gap test. So the fixture
cannot fit, and the harness kills Node first — which is why the result is NODE_FAIL
rather than an output mismatch. It has failed identically on every completed CI run since
the fixture was introduced.

Introduced by 6157e569b — "fix(runtime): stop process.stdin lifecycle calls from killing
input for good (#9676)" (2026-09-04 09:19), landed via merge train #9694. The runtime fix
itself is not in question here; only the fixture's runtime budget is.

Two options, author's call

1. Raise the cap for this fixture — the delays look load-bearing for the race being
   pinned (the comment says the timer is "late enough that the churn role has finished
   collecting first"), so if they must stay, the suite needs a per-test override rather
   than the single PERRY_RUN_TIMEOUT default.
2. Shorten the sleeps — e.g. 2500 ms -> 800 ms, if the ordering they enforce holds at a
   smaller margin.

I did not pick between these because the delays encode timing assumptions about the
stdin-reader race that the fixture exists to pin.

Note for triage

Two other tests print FAIL … PARITY_FAIL in the same job —
test_gap_perfhooks_3088_3008_3010_3011 and test_gap_prop_plan_cache_invalidation — but
both are already recorded in test-parity/gap_snapshot.json, so scripts/gap_snapshot.py
does not count them as regressions and they do not fail the job. They are informational
noise; the fixture above is the only thing making gap-suite (2) red.

https://claude.ai/code/session_01YPfnmWZmSpSWpmnoXvH8z2