[gthyagi]

Submesh injection can lose shared-node values when a parent DOF owner retains no extracted-region cells. Local copies reach another rank's parent ghosts, then owner synchronization overwrites them with untouched values. In the annulus workflow this zeroed four inner-wall topography values before gravity assembly.

Reduce matched values and their multiplicities through the parent field DM before synchronizing owners. Averaging shared contributions preserves replace/add semantics and leaves unmatched parent values unchanged. Production transfer uses distributed PETSc vectors without gathering global coordinates.

The original CI failure exposed a second cause: if an extracted region initially occupies one MPI rank, the filtered DM reports isDistributed() == False. The Mesh constructor then redistributed the child, invalidating rank-local parent matches. Preserve the parent-derived partition when constructing region and surface submeshes; ordinary mesh distribution is unchanged. Surface extraction now checks for a globally empty face stratum, allowing legitimate empty ranks while still rejecting missing/empty surfaces. Independent child repartitioning remains outside coordinate-coincidence injection.

Validation of the updated branch on an isolated Mac PETSc 3.25.0 int32/Open MPI 5.0.10 build:

• Deterministic single-rank-region P1/P2 replace/add cases first fail on the original PR with the same CI errors: replace 5.0, add 4.0. Default-partition controls pass.
• Patched serial suite: 20 passed and one pre-existing expected failure in embedded 1D on-surface evaluation.
• Two MPI ranks: 23 tests passed on each rank, covering injection, region extraction/restriction, starved-rank cell sizes, subsequent mesh construction and surface extraction.
• Four MPI ranks: all 10 focused cases passed on each rank, including empty parent/child ranks, re-extraction, unchanged exterior values and collective rejection of a globally empty surface.
• Style gate, test formatting, Python parsing, diff checks and CI test reachability pass. No new benchmark models were run; the shared Mac installation is unchanged. Fresh Linux/MPICH CI validation is pending.

Earlier Gadi validation applies to the unchanged owner-reduction implementation in feature commit 86cb3cf1, before the partition-preservation follow-up:

• Fresh PETSc int32 build: 18 UW3 tests in serial and on each of 16 MPI ranks; 270 annulus components passed.
• Actual h=1/32 annulus meshes with 1h/2h/3h buffers reproduce the old 2h failure; corrected serial/MPI16 transfers agree within 1.8e-15.
• Six complete Stokes–topography–Poisson regressions restore every saved source value, preserve Stokes fields to roundoff and agree between serial/MPI potential coefficients within 1.01e-11. Audited evidence.

Underworld development team with AI support from Claude Code