CBF flux on driven boundaries, kdtree module identity, global velocity_error

[lmoresi]

Three fixes from the small-issues sweep, with the public adversarial-review round. Closes #407, closes #316, closes #370.

The fixes

#407 — CBF boundary flux was garbage on inhomogeneous Dirichlet walls (2D and 3D). Constrained DOFs don't exist in the global vector, so _assemble_volume_reaction's gather round-trip zero-filled every boundary value before integrating the reaction residual — a g=1 wall read −17.7 for a true unit flux, and every existing test was blind because it measured g=0 walls. DMPlexInsertBoundaryValues now fills the essential values first. Regression tests: exact unit flux through a driven wall (end nodes included), and a corner-semantics test (T=x+y, all four walls driven) that pins the method's inherent corner behaviour — a corner reaction integrates BOTH adjacent walls' fluxes. The review's Couette probe confirms the Stokes case: inhomogeneous velocity walls now give exact (±1, 0) tractions where the old code garbled them; all SolCx/rotated-BC goldens are unchanged (g=0 inserts zeros over zeros), and the rotated-reaction path is untouched by construction.

#316 — order-dependent memprobe failures. A shim kdtree.py shadowed the package attribute on the first import underworld3.kdtree (which test_0000_imports performs), silently dropping the memprobe counters for the rest of the run. Now one module identity via a sys.modules alias; shim deleted; identity regression test. Verified in both import orders.

#370 — the np4 FMG 'tolerance miss' was never a solver problem. velocity_error computed rank-local norms: the same converged solution measured 0.00027 on one rank and 0.00562 on the rank owning the SolCx viscosity jump, and the test asserted per-rank. Now a global reduction (identical on every rank; the seam-DOF weighting approximation is documented). Global np4 error: 0.0006 against the unchanged 5e-3 tolerance; the full test_1017 file passes at np2 and np4. Second instance of the rank-local-diagnostic disease after #351 — worth watching for as a class.

Adversarial review (public, round 1 — posted below)

All three held. The review's probes: Couette all-wall Stokes exactness, mixed natural+Dirichlet bit-identity, golden batteries (test_1018 serial 14 passed, test_1064 np2, test_1065 np2/np4, SolCx goldens), corner-test flake repeats (serial/np2x3/np4x3), both kdtree import orders, pickle/module-name attacks, and an exhaustive collective-safety sweep of every velocity_error call site (the global norm makes it collective — no rank-guarded caller exists). Serial equivalence of the new norm: 1 ulp.

Found on the way, filed separately: #410 — time-dependent Dirichlet BCs NaN out in solve() (upstream PETSC_MIN_REAL insertion reaches mesh.t kernels as −1.8e308; pre-existing, and the reason boundary_flux's time handling is currently moot); #411 — the duplicated compute_*_residual_fields diagnostics still carry the stripped-Dirichlet round trip (READ-27 de-duplication is the right fix).

Verification

• test_1019 (with new tests): serial, np2, np4 — green, repeats clean.
• test_0000+test_0780+test_0101 in both orders: green.
• test_1017 full file: np2 and np4 green.
• Level_1 + tier_a gate: 444 passed, 0 failures.

Manual close needed on merge: #407, #316, #370.

Underworld development team with AI support from Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[lmoresi]

Adversarial review — round 1 (public record)

Independent skeptic pass over the branch diff; live probes serial/np2/np4 in the branch's own build.

#407 — essential-value insert before the reaction residual

• Stokes / mixed-BC corruption attack: HELD. All-wall inhomogeneous Couette (u=(y,0), P2-P1, pressure nullspace): solution exact to 2e-12, recovered tractions exactly (±1, 0) — precisely the case the old code garbled. Adding a natural wall alongside: bit-identical results. Rotated-BC path untouched by construction (reads the stashed solve reaction).
• Golden regression attack: HELD. test_1018 serial (14 passed, incl. SolCx sigma_nn + dynamic-topography goldens), test_1064 np2, test_1065 np2/np4, SolCx goldens — nothing shifted (for g=0 the insert writes zeros over zeros).
• Time-dependent BC attack: could not be realized — and found a pre-existing solver bug instead. No caller passes time= today, and no solve can reach boundary_flux with a live mesh.t Dirichlet wall because solve() itself inserts essential values at PETSC_MIN_REAL — mesh.t kernels see −1.8e308 and the solve NaNs. Filed as Time-dependent Dirichlet BCs are broken in solve(): essential values inserted at PETSC_MIN_REAL, mesh.t kernels see -1.8e308 #410 (with the note that boundary_flux must forward the solve time once that is fixed). The time= plumb in this diff was verified to insert g(t) correctly.
• Scope gap found: the duplicated compute_volume/boundary_residual_fields diagnostics did NOT get the insert (no in-repo callers; READ-27 de-duplication territory). Filed as compute_volume_residual_fields / boundary sibling still strip Dirichlet values (the #407 fix was applied to one of three duplicated variants) #411.
• Corner-semantics test: no flake across serial + np2 x3 + np4 x3.

#316 — kdtree module identity

HELD. All import spellings resolve to the one module, both import orders green, pickle/module-name attacks dead on arrival (KDTree was never picklable; module was always ckdtree). One packaging note, fixed before push: the stale setuptools build cache re-shipped the deleted shim into the local wheel — build/lib.* cleaned; clean checkouts/CI unaffected, and the sys.modules alias preempts a stale file anyway (verified: the identity test passed even with the stale shim present).

#370 — global velocity_error

HELD — the collective-deadlock attack found no victim. The global norm makes the function collective; an exhaustive sweep of every call site (10 test files, src, docs, scripts) found none inside rank guards or rank subsets. Serial equivalence to the old formula: ~1 ulp. Seam-DOF duplication quantified at ~4% of rows at np2 — matches the docstring's disclosed approximation. Bonus: empty ranks now get the finite global value instead of 0/0 NaN.

Verdict

Merge. Follow-ups #410 and #411 filed; neither blocks.

Underworld development team with AI support from Claude Code