test_global_evaluate_after_migration hangs at np=4, and its file is in neither parallel-test glob

[lmoresi]

tests/parallel/test_0760_swarm_cache_migration.py::test_global_evaluate_after_migration
hangs at np=4 on current development (8b7c8b9). It passes at np=2.

Reproduction

mpirun --timeout 90 -n 4 python -m pytest --with-mpi \
    tests/parallel/test_0760_swarm_cache_migration.py::test_global_evaluate_after_migration -q

mpirun reaches its timeout with no test result. The other two tests in the
file pass at np=4:

test	np=2	np=4
test_swarm_cache_valid_after_migration	pass	pass
test_global_evaluate_after_migration	pass	hang
test_global_evaluate_displaced_nodes	pass	pass

Not introduced by anything in flight — reproduced on development itself, and
the same run on bugfix/shrinking-working-set (#609) hangs identically, which is
how it was found.

Why nobody has seen it

scripts/test.sh runs exactly two globs under mpirun:

mpirun -n $PARALLEL_RANKS python -m pytest --with-mpi tests/parallel/test_075*py
mpirun -n $PARALLEL_RANKS python -m pytest --with-mpi tests/parallel/test_10*py

test_0760_... matches neither — test_075* stops at 0759 and test_10* starts
at 1000. The file has never been run in parallel by our own script, at any rank
count. PARALLEL_RANKS also defaults to 0, so ./uw test runs no parallel tests
unless --p is passed.

This is the #570 class exactly: a coverage gap in the glob, hiding a real defect.
The gap is the reason the defect is old and unnoticed rather than new.

Why it matters now

The hanging test is global_evaluate after a swarm migration, which is the
subsystem under active work: #551 (locator anti-scaling), #604 (a located point
returning NaN), #606 (rank-local answers in parallel), and PR #602, which adds a
regression test to this same file and records its validation as
mpirun -n 2 ... 4 passed per rank. At np=2 everything here passes, so that
validation would not have caught this.

Suggested handling

1. Fix the globs so tests/parallel/ is covered by pattern rather than by two
   ranges that happen to have a hole between them — the hole will recur
   otherwise.
2. Run the parallel selection at np=4 as well as np=2. Every test that passes at
   np=2 and hangs at np=4 in this subsystem has been a collective reached by some
   ranks and not others; two ranks is often too few to expose it.
3. Then diagnose the hang itself.

Underworld development team with AI support from Claude Code