wheel: swath1.mps aborts with std::bad_array_new_length after fast MPS parser (#1429)

[ramakrishnap-nv]

Summary

Three wheel tests crash with std::bad_array_new_length on all GPU runners since the fast MPS parser (#1429) merged into main on 2026-07-20:

• tests/linear_programming/test_incumbent_callbacks.py::test_incumbent_get_callback[/mip/swath1.mps]
• tests/linear_programming/test_incumbent_callbacks.py::test_incumbent_get_set_callback[/mip/swath1.mps]
• tests/linear_programming/test_lp_solver.py::test_heuristics_only (also uses swath1.mps)

Symptom

terminate called after throwing an instance of 'std::bad_array_new_length'
  what():  std::bad_array_new_length
Fatal Python error: Aborted

The crash happens inside solver.Solve() before any callback fires.

Key observations

• Fails on all 6 wheel-test runners (rtxpro6000 and l4, amd64 and arm64)
• neos5-free-bound.mps variants of the same tests pass
• conda-python-tests pass on identical GPU configs — the conda binary either parses swath1.mps correctly or solves it at root (0 callbacks → pytest.skip), suggesting a build-configuration difference between conda and wheel binaries
• Tests were previously disabled (May 2026, Disable flaky swath1 incumbent callback tests #1208) for a different swath1 crash (PDLP CUDA graph capture race), then re-enabled in July 2026

Root cause hypothesis

The fast MPS parser (#1429) mis-parses swath1.mps in the wheel build, producing an invalid array size that causes new[] to throw. The conda build is unaffected, likely due to different compiler flags or linked library versions.

Workaround

swath1.mps tests temporarily skipped in #1591 while this is investigated.