feat(evaluate): measured point-location capability + NaN/RBF fallback ladder

[lmoresi]

Implements #392 (stacked on #391 — retarget to development once that merges; the diff shown here is the capability-ladder work only).

What this does

Makes the cell locator's authority (i.e. is it authoritative or an estimator) a measured mesh property instead of an isSimplex() type flag, and completes the fallback ladder agreed on #392 so quad/hex meshes keep working with a fallback that simplex meshes don't need.

Mesh._location_capability() audits face planarity (sagitta) and cell convexity, cached against (_mesh_version, _topology_version) so deform() and adaptation refresh it automatically:

• exact — simplex, manifold, and any quad/hex mesh whose faces are planar and cells convex at machine level. This graduates all valid 2-D quad meshes, deformed included (straight edges are always planar; convexity is equivalent to an untangled mesh — verified 4000/4000 against exact polygon containment on a deformed box) and rectilinear/affine hex boxes to simplex-grade location: the hint bypasses DMLocatePoints.
• continuous — warped hexes within a 5% sagitta tolerance (cubed-sphere class: 64% of faces exactly planar, spherical faces ≤ 1e-2). Authoritative for continuous fields only — a misclassified point lands in the face-adjacent neighbour where continuous FE interpolants agree to O(sagitta × gradient); a face-aligned jump (layered viscosity on a spherical shell) would see O(jump) errors.
• none — badly warped / non-convex cells: no authority.

The ladder: non-authoritative evaluations (capability none, or continuous with a discontinuous variable in the mix) run DMLocatePoints; points it drops (the #390 class — queries exactly on the domain's closed upper faces) are NaN-filled — never zeros — reported via a new unlocated mask, and filled per-variable through rbf_interpolate: bounded, topology-free, honest. NaN survives only if the plumbing is bypassed, which is exactly when it should be visible.

The policy lives in one place (Mesh._hint_is_authoritative) and joins the DMInterpolation cache key (same coords with a different field-continuity mix legitimately need different structures).

Testing

• Three new tier_a tests in test_0503: capability grading of regular/deformed quad and hex meshes; deformed-quad boundary evaluation exact to 1e-8; warped-hex RBF rung (no NaN escapes, no silent zeros, values bounded by data range).
• test_1052 VEP stability: 5/5 (the shear box now takes the authoritative bypass).
• level_1 + tier_a gate: 440 passed.
• Parallel np2 suite: identical to baseline (92 passed; the single failure is the pre-existing test_0765 _deform_mesh-guard issue with its own TODO).

Design record: issue #392 and docs/developer/design/point-location-capability.md.

Underworld development team with AI support from Claude Code

[lmoresi]

CI resolved — the failure was a runner-side flake, not the code. The failing run died during import underworld3 before pytest's first stdout flush (every batch ~2s, zero output, undiagnosable from the log). A faulthandler import smoke step was added to surface the crash point; on the very next run the identical code imported cleanly and the full test job passed (25 min, green). The smoke step is kept as a permanent 5-second guard so any future import-time crash is loud and localised. Local reproduction attempts on both custom PETSc 3.25.0 and conda-forge PETSc 3.25.3 (full CI batch selection via scripts/test.sh) were green throughout.