Strip every launcher's variables before the serial reference child (#675)

[lmoresi]

Summary

Closes #675.

test_parallel_matches_serial_bit_identical spawns a serial reference
subprocess from inside an mpirun, and scrubs the launcher's variables so the
child does not try to join the parent's job. Its scrub list named the Open MPI
family only:

OMPI_*, PMIX_*, PRTE_*, PRTERUN_*, OPAL_PREFIX

MPICH advertises itself through PMI_*. On the Linux CI runner the child
inherited PMI_FD and tried to join a job it was not part of:

[cli_0]: PMIU_write error; fd=9 buf=:cmd=init pmi_version=1
system msg for write_line failure : Bad file descriptor
Abort: internal_Init_thread ... PMI_Init returned -1
assert 15 == 0

macOS runs Open MPI, where that list is complete. So this passed on one
machine and failed on the other, with nothing to do with the operating system —
and it had never been caught because the file matches no glob in either test
script and had never run in CI at all.

The fix already existed thirty lines away

tests/parallel/serial_reference.py, same directory, carries the complete list.
This test hand-rolled a copy and the copy drifted. It now imports the shared
tuple.

The shared tuple gains PRTE_, PRTERUN_ and OPAL_, so the union is strictly
wider than either copy was. Swapping to the shared list unchanged would have
narrowed Open MPI 5 coverage — PRRTE is a separate family from OMPI_, and
the local copy was the only thing naming it.

This is Charter §5 copy-paste-as-reuse, and the divergence between the two
copies is the entire bug.

test_0873 was collateral

test_0873_adapt_collective_stop_mpi::test_premise_the_metric_splits_the_ranks
passes in isolation. It failed only downstream of this abort, which also
produced the rank divergence that hung the batch for 76 minutes — rank 0 in
_from_plexh5(), rank 1 in barrier(). Both files are green together now.

Verification

where	result
Linux CI runner, MPICH, both files, np=2	8 passed in 3.81 s
Open MPI 5.0.10 local, test_0855	4 passed
Open MPI 5.0.10 local, test_1069 (uses serial_reference directly)	7 passed, 1 skipped

The Linux run was obtained with the supervisor from #678, which turned 76
minutes of silence into a named diagnosis in six.

Not fixed here

The coverage hole that hid this — three parallel files matching no glob in
either script — is #615's subject. With this merged, that PR's blocker is gone.

Underworld development team with AI support from Claude Code

🤖 Generated with Claude Code

https://claude.ai/code/session_01E87Q7KrpapxeQiLD1RiNXv

[Copilot]

🟢 Approval recommended

The change is narrowly scoped to test subprocess environment scrubbing and correctly centralizes the launcher-variable prefix list to prevent MPICH/Open MPI divergence.

Pull request overview

This PR fixes a cross-MPI-launcher environment inheritance bug in test_parallel_matches_serial_bit_identical by centralizing the “strip launcher variables” logic, ensuring a serial reference subprocess spawned from within an MPI job does not attempt to join the parent MPI world (notably under MPICH via PMI_*).

Changes:

• Replaces the hand-rolled Open MPI–only env-var scrub list in test_0855 with an import of the shared _MPI_ENV_PREFIXES tuple from serial_reference.py.
• Expands the shared scrub-prefix tuple to include Open MPI 5 runtime families (PRTE_, PRTERUN_, OPAL_) in addition to existing prefixes (including PMI_* for MPICH).

File summaries

File	Description
tests/parallel/test_0855_mesh_smoothing_parallel.py	Uses the shared MPI launcher env scrub tuple to reliably spawn a serial reference subprocess under multiple MPI implementations.
tests/parallel/serial_reference.py	Broadens and documents the shared MPI launcher env scrub tuple to cover additional Open MPI runtime prefixes.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 1
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.