[lachlangrose]

Draft, stacked on #366. This branch contains #366's commit, so its diff against master shows that until #366 merges. Merge #366 first.

Step 2 of consolidating the structured-grid geometry code: BaseStructuredSupport now extends StructuredGrid3DGeometry instead of re-implementing the same geometry.

Change

• BaseStructuredSupport(StructuredGrid3DGeometry, BaseSupport) keeps only the support-specific parts: the constructor (takes cells, the geometry class takes nodes), the onGeometryChange hook, the interpolator link, position_to_cell_corners / _vertices, node_indexes_to_position, element_size, vtk, and its more tolerant check_position. About 330 net lines removed from _3d_base_structured.py.
• The geometry setters (nsteps, step_vector, origin, maximum) call a no-op _on_geometry_change() hook; the support maps it to onGeometryChange(), so setter behaviour is unchanged.
• BoundingBox.structured_grid imports loop_common.supports lazily: supports now imports geometry, so the module-level import would be circular.
• neighbour_global_indexes is removed from StructuredGrid3DGeometry. It is a finite-difference stencil helper that StructuredGrid already defines for itself, and inheriting it would have added a method that does not apply to TetMesh / P2TetMesh.

How I checked behaviour is unchanged

test_structured_support_reference.py records 611 arrays from StructuredGrid, TetMesh and P2TetMesh on three grids (geometry, nodes/elements, point location including exact corner and face points, value/gradient evaluation, regularisation operators, each setter, onGeometryChange call counts, constructor validation) and compares them with tests/data/structured_support_reference.npz (316 KB). The reference was generated from master before this change, so it pins the original behaviour. Regenerate with python packages/loop_common/tests/test_structured_support_reference.py --update.

Mutation checks: dropping the setter hook, making inside() exclude the boundary, and removing the cell-index clip each make the test fail.

• packages/loop_common/tests: 183 passed
• packages/loop_interpolation/tests: 606 passed, 46 skipped, 2 xfailed (includes the golden tests)
• tests: 730 passed, 6 skipped
• ruff: no new findings

Known and not addressed

• Giving step_vector as integers makes the nsteps setter raise (_step_vector /= change_factor cannot cast to int). Unchanged: behaves the same before and after, and is not pinned by the reference.
• Not done: the 2D grid, RectilinearGrid, and moving the remaining support members up.
• Not checked: whether the QGIS plugin or map2loop subclass or access BaseStructuredSupport internals.

🤖 Generated with Claude Code