Third failure on this feature, and this one broke everything rather than just
the diagnostic: 43 failures across every run-tests.cjs invocation.
Error: EINVAL: invalid argument, fsync
Emitted 'error' event on WriteStream instance
Node opens a WriteStream for a --test-reporter-destination and FSYNCS it on
close. fsync on /dev/null is EINVAL — it is a character device, not a regular
file. So os.devNull is not a usable reporter destination at all, and every
chunk crashed on exit.
The sink only ever needed to be a regular file that stays empty, since the
reporter writes its real output through appendFileSync to the path in
GSD_RUN_TESTS_EVENTS_FILE. It is now one fixed file inside the existing events
dir, pre-created rather than relying on the stream's create-on-open, and left
empty by design. One sink for the whole run, not per chunk, so its path length
stays constant — reporterOverhead feeds FIXED_OVERHEAD, which is computed once
before chunking, and a variable-length path would silently mis-account the
Windows argv ceiling.
The comment that named devNull now says why the destination must be a regular
file, so this does not get re-optimized back into the same crash.
Reaching for devNull was the mistake: it looks like the obviously correct way
to discard output, and it is, for a pipe or an fd — but not for something Node
is going to fsync. Each of the three failures on this feature was a different
edge of the same assumption, that a reporter destination behaves like ordinary
output.
The regression test asserts the closest externally observable consequence — a
normal run must not surface the EINVAL/fsync text. The argv construction lives
inside main() with no exported seam, and the sink is swept before a test could
stat it; adding a seam purely to assert that is left out rather than reshaping
production code for the test. Stated plainly rather than implied.
The failing T1 is untouched and still red.
Verification runs on the remote runner.
Refs #4012
7f2af28639