Tet mesh quality: one validity criterion, Delaunay refinement, ODT, Klingner–Shewchuk, tet10

The end-to-end demo met a raw-CDT sliver that the validity gate accepted (exactly positive) and
the solver refused. Now:

- One validity criterion, `fieldcraft_model::admissible_jacobian` (normalized Jacobian above
  1e-12), used by the implicit solver (fem `positive_inverse`), the explicit tet4/hex8
  kinematics and the mesh gate, which proves it on the whole element (tet4: the solvers' own
  Jacobian; curved kinds: a Bernstein floor), so whatever passes the gate passes the solvers.
- CDT rounding repaired by flips on the rounded mesh (Diazzi et al. 2023 §5.1) instead of the
  ±2-ulp nudge.
- `fieldcraft-mesh::tet`: a mutable tetrahedral complex whose every change is a checked
  replacement (positive and admissible tetrahedra, faces matched with opposite orientation) with
  an exact undo log; edge removal (Klincsek), multi-face removal, Delaunay insertion with
  boundary faces, edge collapse; a curvature- and thickness-graded size field; constrained
  Delaunay refinement with encroachment splits; ODT smoothing; the Klingner–Shewchuk schedule;
  boundary vertices moved onto the exact B-rep surfaces at the end; tet10 with mid-edge nodes on
  the geometry.
- `mesh.tet.generate` gains size, order, quality, quality_target, curvature_angle, grading,
  max_passes, max_points and reports the CDT's and the mesh's quality; `solve.probe` takes a set
  and reports its largest von Mises stress.
- Oracle cases filleted-bracket-fillet-stress(-coarse): the fillet's peak von Mises stress on
  tet10 meshes of 2k and 13.5k elements, 213.3 MPa (a four-level study stays within 1.8%); the
  CDT cases keep the raw tetrahedralization.
- Corpus quality columns: all 20 valid parts mesh and pass the gate.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NmAttHhUZTokgPeEJia8sm

docs/architecture.md

@@ -65,9 +65,13 @@ Rules beyond the layer order:
    perturbation for degeneracies), boundary recovery by the robust constrained Delaunay method of
    Diazzi et al. 2023 (implicit Steiner points on indirect predicates, gift-wrapping facet
    recovery; `fieldcraft-mesh::cdt`, command `mesh.tet.generate`), so every boundary triangle and
-   its B-rep face tag survives. Then constrained Delaunay refinement to a
-   size field, ODT smoothing interleaved as in Tournois et al. 2009, and the Klingner–Shewchuk 2007
-   improvement schedule against slivers.
+   its B-rep face tag survives; rounding is repaired by flips (§5.1). Then the quality stage
+   (`fieldcraft-mesh::tet`): a curvature-graded size field, edge-collapse coarsening, constrained
+   Delaunay refinement with boundary points projected onto the B-rep faces and curves (exact
+   surfaces from SolidCraft where analytic, the tessellation otherwise), ODT smoothing
+   interleaved as in Tournois et al. 2009, the Klingner–Shewchuk 2007 improvement schedule
+   against slivers, and tet10 with mid-edge nodes on the geometry. Every step is a checked local
+   replacement, so the mesh stays valid throughout.
 4. **Hexahedral meshing (the core).** Grid-based all-hex first, the one family that meshes any
    input validly. `fieldcraft-mesh::octree` (command `mesh.hex.octree`): a 27-tree over the part,
    refined by curvature, thickness and narrow faces, balanced and cut by the 3-refinement
@@ -107,7 +111,10 @@ Rules beyond the layer order:
    faces, edges and vertices become geometry tags that sets select. It meshes the analytic
    shapes of the verification cases.
 6. **Quality.** Scaled Jacobian, aspect ratio, dihedral angles; every generated mesh is checked
-   for positive Jacobians and conserved volume before it reaches the model.
+   for valid Jacobians and conserved volume before it reaches the model. Validity is one
+   criterion shared with both solvers (`fieldcraft_model::admissible_jacobian`: normalized
+   Jacobian above 1e-12): the gate proves it on the whole element, the solvers check it at their
+   integration points, so a mesh the gate accepts is never refused by a solver.
 
 ## Model definition