node-opcua / node-opcua / 30173685538

92%

push

github

ci: capture the real Windows exit code instead of a truncated one

The Windows test step ran under bash, so $? was a POSIX wait status and
only kept 8 bits. That truncation is what made the recurring failure
unreadable: an NTSTATUS ending in 0x7F (native crash) and a genuine
exit(127) collapse to the same "127", and the step then asserted a
native/loader-thread crash it had no way to distinguish.

Every abnormal exit on record (5 runs, Jun 30 - Jul 6) came from
build_on_windows (24.15.0), at a different test each time, ~1s after the
last line of output with no mocha summary. Linux and Windows 22.x never
reproduced it - in run 28754712923 the 22.x job exited 1 on a real test
failure while 24.15.0 exited 127 on the same commit.

- run the step under pwsh: $LASTEXITCODE keeps all 32 bits, and the code
  is now logged in hex too (0xC0000005 reads very differently from 0x7F);
- enable --report-on-fatalerror / --report-uncaught-exception and upload
  the reports on failure. stdout to a pipe is async on Windows, so the
  final writes are lost when the process dies; the diagnostic JSON is the
  only evidence that survives;
- reword the abnormal-exit error: missing stdout does not prove a native
  crash, which is what the old message claimed;
- do not pin the ubuntu matrix to 24.15.0. Linux has never shown this
  crash, so pinning there would only cost coverage of current Node 24.

Also drop a stray bare `_` statement in checkMemoryConsumption(), a
ReferenceError on every call - latent only because that path needs
NODEOPCUATESTREPORTER, which CI does not set.

This is instrumentation, not a fix: the root cause is still unknown.

19983 of 23297 branches covered (85.77%)

169561 of 183922 relevant lines covered (92.19%)

457630.9 hits per line