Provide a multigrid hierarchy for the constrained-traction regression

[gthyagi]

The constrained-traction SolCx regression times out on Gadi with PETSc 3.25.4 using both 32-bit and 64-bit indices. Its single-level 32×32 mesh causes the high-contrast augmented system to fall back to GAMG. A diagnostic recorded 21,666 velocity iterations in 90 seconds; the unchanged upstream test exceeded its 180-second timeout.

Build a 16×16 base mesh and refine once, preserving the 32×32 fine grid while providing geometric multigrid. The hierarchy allows the existing test to finish and check the augmented traction, bare-multiplier negative control and topography against SolCx. Physical parameters, tolerances, accuracy bounds and production solver defaults are unchanged.

Validation

• Gadi PETSc 3.25.4 int32 and int64 container checks pass with the hierarchy, taking approximately 35–40 seconds in the targeted probes.
• A separate upstream-source baseline check passes in 32.72 seconds with only this mesh change applied, establishing that it does not depend on the integer-width fixes in PR Fix PETSc integer-width mismatches in boundary registration and JIT callbacks #813.
• Paired verification job 180509622 passed all 23 selected serial tests for each integer width, including this regression, plus the selected two- and four-rank MPI checks.
• Python syntax and diff whitespace checks pass. No shared native Pixi installation was rebuilt for this validation.

This change is separate from PR #813's PETSc integer-width fixes and from the benchmark-only 3D DFZ memory changes.

[lmoresi]

Issues in this branch's territory

Territory is tests/test_1063_constrained_traction.py.

issue	verdict
none open against this test	—
#801 test_1063_constrained_freeslip_parallel[ti]: topography 37.75 vs 37.08 at np=4	still live — a different test in the same family, and a value discrepancy rather than a timeout. Nothing here touches it.

Beyond the territory: this unblocks #744

PR #744 (the discrete adjoint) is red on exactly this test:

FAILED tests/test_1063_constrained_traction.py::test_constrained_topography_carries_the_augmentation_share

scripts/triage.py (#820) classifies that failure as live — the test is on development, so #744 is not stale in the way #785/#789/#795/#800 are. But this PR is the reason the test times out: a single-level 32×32 mesh drops the high-contrast augmented system onto GAMG at 21,666 velocity iterations in 90 s. Giving it a 16×16 base refined once keeps the 32×32 fine grid and provides the hierarchy.

So this should merge before #744 is re-run, and #744's CI is then worth re-checking before anyone treats its failure as its own.

Underworld development team with AI support from Claude Code

[lmoresi]

Approved for merge. Adjudication posted above; the multigrid hierarchy is confined to the test fixture and the physical parameters are unchanged, so the assertions still gate what they gated. Merging this ahead of #744, whose CI failure is this test timing out.