Repository navigation
test: start-wiring's real-child poll is a 2 s budget, and flakes under full-suite load #369
Description
Activity
A second instance of the same class, seen while gating #371 (which touches no receiver file).
receiver/test/start.test.mjs:433, "the triggers watch ARMS under test, and a shut-down watch writes NOTHING
(issue #301)", failed once in two consecutive full CI-posture runs and passed 3/3 in isolation (20/20 for the
file). The assertion is line 473:a live watch reloads, or this test is asserting silence from a watch that never worked
Worth noting because this one is already the mitigated version: it POLLS on a 5 s deadline rather than
sleeping, and its own comment says why ("a fixed wait is a fuse on a loaded runner"). It still missed. So the
variable isfs.watchdelivery latency plus the 150 ms debounce under a fully loaded runner, and 5 s is not
always enough on a machine running the whole suite in parallel.Same shape as the
start-wiring.test.mjspoll above, same consequence: a red build on a tree nobody touched,
inside a required check. Filed here rather than as its own issue because the fix is one decision, not two:
what deadline do real-time polls in this suite get when the whole suite is running?Both instances pass in isolation, so neither is evidence of a product defect.
- added a commit that references this issue
on Sep 22, 2026
Found by the gate on #368 (issues #348/#352), out of scope for that PR and filed rather than fixed there.
worker/test/start-wiring.test.mjs:1923, "a real err.reason token rides the hook's argv end to end, through a REALscript (review finding)", polls for a genuinely spawned
shchild to append a file:That is a 2 s budget. Under the full parallel suite it can be missed: an executing reviewer saw it fail once in two
full CI-posture runs on an otherwise clean tree, and pass 3/3 in isolation (91/91 for the file). The assertion that
fails is
the hook really spawned, which reads like a real defect in the failure hook and is not one.This matters more than an ordinary flake because the suite runs in a REQUIRED check (
contract-testsruns it threetimes per job: plain, under
test-count-check.mjs, and on a clock shifted 399 days). A one-in-N failure there is ared build on a tree nobody touched, which is the shape issue #284 already cost this project a day of merges to.
Suggested fix, and it costs nothing on the happy path because the loop exits on first existence: raise the bound
(for example
i < 500, a 10 s ceiling). Worth checking whether any sibling in the file polls a real child on thesame 2 s budget and raising those together rather than one at a time.
Not a product defect: nothing about the hook, its argv or its env threading is in question, and the test proves
what it says it proves when it is given time to.