meta(grids): non-box domains and richer boundary conditions — geometry seam, mask, Robin, tensor diffusion

[kylebeggs]

Umbrella for the four issues that extend the package beyond "axis-aligned box, scalar κ, Dirichlet/Neumann/Periodic". The comment threads already tie them together: #61's binary mask is the φ ∈ {0,1} special case of #69's phase-field candidate; Robin on an irregular domain is #62 ∩ #69; #70's face-averaging policy depends on whether the coefficient Diffusion carries becomes φD.

Sub-issues, in order

1. feat(grids): finite differences on non-trivial geometry — pick the seam #69 — pick the geometry seam (staircase mask / Shortley–Weller / generalized FD / level-set / phase-field). An investigation with a convergence study, not a build. Its answer decides the shape of feat(grids): active-cell mask so stencil operators can skip inactive cells #61: an active::Field{Bool} and a φ::Field{T} want to be the same seam.
2. feat(grids): active-cell mask so stencil operators can skip inactive cells #61 — active-cell mask on CartesianGrid. Unblocks the downstream microvascular model; per-cell operator_diagonal has a closed form under mirroring.
3. feat(boundaries): Robin BC (a·u + b·∂u/∂n = c) #62 — Robin BC on the six box faces. Self-contained; the per-face Field coefficient question is the real design call. Prior art for the irregular-domain version is shared with feat(grids): finite differences on non-trivial geometry — pick the seam #69 option 4.
4. feat(operators): anisotropic diffusion ∇·(D∇u) with a full tensor D #70 — anisotropic tensor diffusion. Stage 1 (tensor_field, lift the two scalar-eltype guards) can proceed at any time; stages 2–3 (tensor face averaging, cross-derivative terms) wait on feat(grids): finite differences on non-trivial geometry — pick the seam #69 because the averaging policy for φD across a geometric interface differs from D across a fiber-orientation jump.

Invariants that bind every sub-issue

• Leaf bodies stay array-level; a per-cell variable stencil is the thing to avoid, and an index-driven gather is the way to avoid it.
• apply_adjoint! remains the exact discrete transpose, checked by ⟨Lx,y⟩ = ⟨x,Lᵀy⟩ on a materialized small problem. Corner ghosts on forests and slabs are garbage by design today; a cross-derivative leaf needs the corner-aware exchange before it touches those paths.
• operator_diagonal becomes per-cell; multigrid depends on it.
• Uniform CartesianGrid first for every item. Masked or tensor BlockForest / slabs are follow-ups.

---

_{🤖 Beep boop — Claude filed the umbrella; Kyle is not standing under it talking to himself.}