[lachlangrose]

Summary

This pull request implements stage S1 of spec 005 (coordinate and fault-frame correctness). The spec is in specs/005-coordinate-frame-correctness/. All 17 S1 tasks are done. Stages S2 to S6 are not planned yet (OP1).

Modelling observations, reference geometry and public queries stay in world coordinates. Local assembly uses the computational frame. Each boundary converts points and directions one time only.

Changes

• Rigid transforms (loop_common.geometry): new validate_rigid_rotation and RigidRotationError. BoundingBox.set_local_transform and StructuredGrid3DGeometry.rotation_xy reject scale, shear, reflection and non-finite matrices, and keep the prior state. New BoundingBox.local_enclosure.
• Support resize: the origin and maximum setters stored a cell count as the node count, so the support lost one cell. They now enclose the requested bounds (FR-014).
• Interpolator: new public world_to_local_points, local_to_world_points, world_to_local_vectors and local_to_world_vectors methods.
• Reference constraints: orthogonal, equality and inequality references are sampled at world points, and the rows are installed in the target-local basis. Invalid rows are counted in builder.excluded_rows. The support uses the local enclosure of the world box, and the restored-point expansion works in world space.
• Fault frame:
  • An explicit fault_center is a world point; the factory does not project it.
  • An inferred centre keeps its elevation.
  • The slip is made orthogonal to the normal, and a slip parallel to the normal raises ValueError.
  • One extent-direction row is added at the centre.
  • The requested fault_buffer and force_mesh_geometry are kept; the factory no longer replaces them with the default 0.2 buffer.
  • fault_ellipsoid uses the world centre.
  • The caller's normal array is not changed.
• COMPAT.md: a new migration notice.

Golden change

packages/loop_common/tests/data/structured_support_reference.npz: 27 resize values change (after_origin and after_maximum nsteps and maximum). The old values did not enclose the requested bounds. All other keys keep their old values. The full record is in verification.md ("Golden review").

Test plan

• ruff check LoopStructural packages: pass
• pytest (root): 1206 passed, 5 skipped, in 867 s. Before S1: 755 passed, 5 skipped.
• pytest packages/loop_common/tests packages/loop_interpolation/tests: 912 passed, 47 skipped, 2 xfailed
• New analytic tests: test_rigid_transform_contract.py (42), test_reference_constraint_frames.py (312), test_fault_frame_analytic.py (55), test_fault_buffer_bounds.py (84). They take about 5 minutes.
• Before S1 (commit ebc7c699), the new tests fail as expected, and the rollback tests pass. See verification.md.
• Reviewer: decide where the R2/R5 matrix runs in CI (OP2). It is in the default pytest run now.

Notes

• black --check reports 2 changed files that are not Black-clean: _fault_segment.py and _discrete_interpolator.py. Neither file was Black-clean before this change, and only the task lines were changed.
• 7 minor review findings stay open, most of them weak test assertions. They are listed in specs/005-coordinate-frame-correctness/run-state.json.
• builder.build() ignores the top-level rtol, atol and maxiter arguments and gives no warning (they must go in solver_kwargs). This is not changed here.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XADRvJLjMM6vnWTMwVqi61