Summary
Found while unskipping the test_0502 MPI tests for #326 (BF-13): on a BoxInternalBoundary mesh at np2, |∫ n_y dS| over the internal surface = 0.9375 instead of 1.0 — exactly one seam facet's normal is flipped (1 − 2/32). Scalar integrands are exact; only VECTOR integrands involving the internal-facet normal are affected, and only at partition seams.
Status
The two affected tests (internal_normal_ny, internal_normal_weighted in tests/test_0502) are re-skipped with an accurate reason and a TODO(BUG) marker at the site. The other four MPI tests in that file are now unskipped and green at np2 after the BF-13 fix (#326).
Impact
Any np>1 surface integral of a vector quantity over an internal boundary (fluxes across internal interfaces, internal traction integrals) can be silently wrong by O(seam-facet count / total facets).
Related: 2026-07 audit loose-ends/parallel family; discovered during PR #326 gating.
Underworld development team with AI support from Claude Code
Summary
Found while unskipping the test_0502 MPI tests for #326 (BF-13): on a BoxInternalBoundary mesh at np2, |∫ n_y dS| over the internal surface = 0.9375 instead of 1.0 — exactly one seam facet's normal is flipped (1 − 2/32). Scalar integrands are exact; only VECTOR integrands involving the internal-facet normal are affected, and only at partition seams.
Status
The two affected tests (internal_normal_ny, internal_normal_weighted in tests/test_0502) are re-skipped with an accurate reason and a TODO(BUG) marker at the site. The other four MPI tests in that file are now unskipped and green at np2 after the BF-13 fix (#326).
Impact
Any np>1 surface integral of a vector quantity over an internal boundary (fluxes across internal interfaces, internal traction integrals) can be silently wrong by O(seam-facet count / total facets).
Related: 2026-07 audit loose-ends/parallel family; discovered during PR #326 gating.
Underworld development team with AI support from Claude Code