BoundedCommandTests and BoundedProcessTests failed nondeterministically:
4/20 runs idle, 7/8 under CPU saturation. Three separate causes.
1. #require inside a retry loop. waitForPID polled through readPID with
`try?`, but #require records an issue even when its error is swallowed,
so the first read of a created-but-not-yet-written pid file failed the
test outright. Added a non-recording pollPID for the polling path and
kept the recording read as the authoritative final attempt. The two
single-read call sites now poll too - echo $$ > file creates and writes
in two steps, so any single read can see a missing or empty file.
2. A 0.1s deadline racing process spawn. BoundedCommand starts its timeout
concurrently with the spawn, so the deadline also bounded /bin/sh
starting and publishing its pid; under load the child was killed before
it ever wrote the file. Wait for the pid while the run is in flight and
give the child a deadline well clear of spawn cost.
3. A 1s per-process budget on the concurrent fan-outs. Instrumenting the
deadline showed a stalled run observed all 64 exits at ~3.1s, clustered
within 100ms of each other - a global stall, not a straggler. Those
tests assert that no exit is lost during monitor registration, not
latency, so the timeout should not double as a performance assertion.
A missed exit still fails: the 50ms pollUntilExit fallback would never
complete.
Also widened waitUntilGone, since reaping is asynchronous.
No production code changed. Proof: 30/30 idle and 20/20 under full 32-core
saturation, against 16/20 and 1/8 before.
Peter Steinberger