Use uw.kdtree.KDTree in mesh.adapt (drop scipy.spatial.cKDTree)

[lmoresi]

Summary

• Three sites in `discretisation_mesh.py` used `scipy.spatial.cKDTree` for the mesh-to-mesh variable transfer that runs during `Mesh.adapt`. These were the only `scipy.spatial.cKDTree` usages left in `src/`.
• Replaced with `uw.kdtree.KDTree` (nanoflann-backed) — the same backend already used by `read_timestep`, swarm migration, and the RBF interpolator path.
• `.query` API is scipy-compatible — same `(d, i)` return, same `k=` — but defaults to squared distances. Passing `sqr_dists=False` preserves the previous linear-distance semantics so the existing threshold (`1.0e-10` for exact-match) and IDW weighting (`1.0 / (dists + 1e-30)`) are unchanged.

Sites changed

• `sub_vertex_map` (line 1247) — parent-submesh vertex match.
• `_re_extract_from_parent` variable IDW transfer (line 1385) — kept k=3 (2D) / k=4 (3D) IDW.
• `_build_dof_map` (line 1440) — parent-submesh DOF match.

Test plan

• No `scipy.spatial.cKDTree` references remain in `src/` (verified by grep).
• `pytest -m "level_1 and tier_a"` on `amr-dev` — 56 passed, 3 skipped, 0 failed.

Notes

This is a minor consistency fix. The bigger architectural item — extracting the global point-location toolkit from `global_evaluate_nd` into a reusable swarm-shaped utility so `mesh.adapt` and `read_timestep` can stop rebuilding rank-local trees redundantly — is recorded in the planning file and is not part of this PR.

Underworld development team with AI support from Claude Code

[Copilot]

Pull request overview

This PR removes the remaining scipy.spatial.cKDTree usage in Mesh.adapt’s mesh-to-mesh transfer paths by switching to uw.kdtree.KDTree (nanoflann-backed), aligning adapt-time behavior with other parts of the codebase already using the UW KD-tree backend.

Changes:

• Replaced parent↔submesh vertex matching KD-tree from scipy.spatial.cKDTree to uw.kdtree.KDTree.
• Replaced adapt-time variable IDW transfer KD-tree from scipy.spatial.cKDTree to uw.kdtree.KDTree while preserving linear-distance semantics via sqr_dists=False.
• Replaced parent↔submesh DOF matching KD-tree from scipy.spatial.cKDTree to uw.kdtree.KDTree.

Comments suppressed due to low confidence (1)

src/underworld3/discretisation/discretisation_mesh.py:1442

• Same as _build_vertex_map: KDTree.query() leaves dists as shape (n, 1) for k==1, so matched becomes 2D and indices[matched] will raise (since indices is 1D). Flatten dists to 1D (or update the query call/implementation to return 1D dists when k==1) before using it as a mask.

        tree = uw.kdtree.KDTree(sub_var.coords)
        dists, indices = tree.query(parent_var.coords, sqr_dists=False)
        matched = dists < 1.0e-10

        # indices[matched] maps parent row → sub row
        parent_rows = np.where(matched)[0]
        sub_rows = indices[matched]

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

uw.kdtree.KDTree.query() returns i reshaped to 1D for k==1, but it leaves d as shape (n, 1). Here that makes matched 2D and then indices[matched] will raise because indices is 1D. Flatten dists (e.g., dists = numpy.asarray(dists).reshape(-1) / dists = dists[:,0]) before building the boolean mask, or adjust the query call/implementation so dists is 1D when k==1 (SciPy-compatible).

Suggested change

	dists, indices = tree.query(self.parent.X.coords, sqr_dists=False)
	dists, indices = tree.query(self.parent.X.coords, sqr_dists=False)
	dists = numpy.asarray(dists).reshape(-1)

[Copilot]

In the IDW transfer path, old_coords is stored via numpy.array(var.coords) (strips UnitAwareArray metadata), but new_coords = old_var.coords can be unit-aware when model units are active. uw.kdtree.KDTree.query() will raise when the tree was built with dimensionless points but query coords have units. Consider using .coords_nd for both arrays (preferred for internal algorithms) or otherwise ensure both old_coords and new_coords are consistently unitless/unitful before building/querying the tree.