test_1052 VEP stability tests fail under PETSc 3.25 (variable-dt pure-VE blow-up), pass in level_1 gate

[lmoresi]

Summary

The three level_2, tier_a tests in tests/test_1052_VEP_stability_regression.py fail
in a local amr-dev worktree running PETSc 3.25.0 (numpy 2.4.6). They are not in
the level_1 and tier_a quick gate, so CI on development stays green — but they are
tier_a (hardened) and should pass.

Failing tests:

• test_pure_ve_variable_dt_accuracy — max|err| = 240.24 vs threshold 0.10
  (pure viscoelastic, variable timestep; a blow-up, not drift)
• test_vep_yield_lock_fixed_dt
• test_vep_snes_no_divergence_loading_through_yield

This is environment/PETSc, not code

• The failure is byte-identical (max|err| = 240.24221587183746) on both the current
  development constitutive_models.py and a feature branch — swapping the source does
  not change it.
• It persists through a clean build/ wipe (rm -rf build/lib.* build/bdist.* +
  rebuild), so it is not a stale-build artifact.
• The pure-VE test returns before any yield / _combine_yield path, so recent yield-law
  work cannot be the cause.

Taken together: the failing behaviour is intrinsic to this worktree's environment
(PETSc 3.25.0), not to any code change on development.

Reproduce

pixi run -e amr-dev python -m pytest \
  "tests/test_1052_VEP_stability_regression.py::test_pure_ve_variable_dt_accuracy" -q
# -> AssertionError: max|err|=240.2422 (threshold 0.10; current head ~0.06)

Hypothesis / suggested investigation

• Most likely a PETSc 3.24 → 3.25 behaviour change reaching the variable-dt VEP path
  (a SNES/line-search/TS default, or a constants[] / DS interaction), rather than a UW3
  regression. reference_env_numpy25_petsc4py_gmsh notes a PETSc3.24→3.25 uw-rebuild.
• Confirm by running the same test against a PETSc 3.24 build; if it passes there, pin the
  changed PETSc default.
• If it is a genuine variable-dt VEP accuracy regression, that is worth fixing on its own
  merits; if it is an env/threshold sensitivity, the test may need an env guard or a
  documented tolerance.

Notes

Discovered while root-causing an unrelated consistent-Newton Drucker–Prager Jacobian
recursion (fixed separately). These VEP-stability failures are pre-existing and orthogonal
to that work; filing for a dedicated session.

[lmoresi]

Root-caused — and the environment/PETSc hypothesis in this issue turned out to be wrong. Fix in #391.

What actually happens: uw.function.evaluate returns silent zeros for query points lying exactly on the domain's closed upper boundary faces of quad/hex meshes (top face of the shear box included). PETSc's point location half-open convention drops those points; the evaluator zero-fills dropped points. The semi-Lagrangian stress-history trace-back slides departure points along the top boundary, so every step reads ψ* = 0 at top-boundary nodes, and the BDF history update amplifies that corruption exponentially (~×1.3/step) — the max|err| = 240 blow-up. Velocity stays machine-exact throughout; the solver was never at fault (an exact-LU run reproduces the blow-up identically).

Why it's not PETSc 3.25: PETSc's plexgeometry.c has only cosmetic changes between v3.24.2 and v3.25.0. The regression is UW3 commit 17a5a8d (May 24): the kd-tree cell hint used to be passed to DMInterpolationSetUp_UW unconditionally, prefill­ing the dropped-point recovery map; when the hint pointer became the DMLocatePoints-bypass switch (simplex/manifold only), quad/hex meshes silently lost the recovery. The "byte-identical failure across branches" observed here was consistent with that — both branches tested postdate May 24. test_1052 was added May 4 and recorded ~0.06; nothing in the level_1 gate evaluates at non-nodal boundary points on a quad mesh, so this went unnoticed.

Fix (#391): an explicit hintAuthoritative flag separates the hint's two roles — bypass (simplex/manifold, unchanged) vs dropped-point recovery (quad/hex, restored). After the fix the pure-VE run gives max|err| = 0.063, matching the test docstring, and all 5 tests in test_1052 pass. A level_1/tier_a regression test now evaluates a linear field on every closed boundary face of quad and simplex boxes.