Fix global_evaluate parallel hang, shape mismatch, and OOB classification

[lmoresi]

Summary

Fixes #113 — global_evaluate shape mismatch and parallel hang in DDt semi-Lagrangian.

Three related issues addressed:

• Parallel hang in evaluate_nd: petsc_interpolate had early returns for empty coords that skipped DMLocatePoints — a collective PETSc operation on the mesh DM communicator. When some ranks had zero interior points (all out-of-domain), they returned early while others blocked in the collective. Fixed by removing the early returns and ensuring all ranks participate.

• Shape mismatch in global_evaluate_nd: Return arrays were sized to the post-migration particle count rather than the original input count. Now pre-allocated with NaN so the shape is always correct — any lost points remain NaN rather than causing a shape error or returning uninitialised data.

• points_in_domain misclassifying extreme OOB points: The boundary control-point sign heuristic had no distance ceiling, so points very far from the domain (e.g. 1e12) could be classified as "interior". Added a domain bounding radius check. Also set ignoreOutsideDomain=True in DMInterpolationSetUp_UW so PETSc silently skips unlocatable points rather than crashing.

Limitations

This fix prevents crashes and ensures correct array shapes, but does not fully resolve AdvDiffusion accuracy on geographic meshes. The RegionalGeographicBox mesh still lacks lateral return_coords_to_bounds (only radial clamping is implemented), so DDt upstream points that advect sideways out of the regional box will be RBF-extrapolated rather than clamped. The extrapolated values are flagged in the mask but may not produce good advection results — lateral boundary clamping for geographic meshes is a separate issue.

Test plan

• Serial: global_evaluate with in-domain and OOB points returns correct shape and mask
• Parallel (4 ranks): no hang, correct shape, OOB points flagged as exterior
• Parallel (8 ranks): same
• Extreme OOB points (1e12): no crash, correctly classified as exterior
• Tier A level 1 tests: 33 passed, 0 failures

Underworld development team with AI support from Claude Code

[Copilot]

Pull request overview

Fixes global_evaluate/evaluate_nd parallel robustness issues by ensuring collective PETSc calls are not skipped on ranks with zero “interior” points, returns are always sized to the original input coordinate count, and extreme out-of-bounds points are not misclassified as interior.

Changes:

• Pre-allocate global_evaluate_nd() return arrays to the original input size (fill with NaN / extrapolated mask defaults) to avoid shape mismatches after swarm migration.
• Remove petsc_interpolate() early-returns for empty coordinate arrays to prevent MPI deadlocks around collective PETSc operations.
• Add an extreme-OOB distance ceiling to Mesh.points_in_domain() and enable ignoreOutsideDomain=True during DMInterpolationSetUp_UW.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
src/underworld3/function/_function.pyx	Fixes global_evaluate_nd output sizing and removes empty-coords early returns in PETSc interpolation path; adds an unconditional update_lvec() call.
src/underworld3/function/_dminterp_wrapper.pyx	Ensures DMInterpolationSetUp_UW runs collectively even with zero points and enables ignoreOutsideDomain to avoid crashes on unlocatable points.
src/underworld3/discretisation/discretisation_mesh.py	Adds a domain bounding-radius check to prevent extreme-OOB points being classified as interior by the boundary-control-point heuristic.

---

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

[Copilot]

mesh.update_lvec() is now called unconditionally in evaluate_nd() before petsc_interpolate(), but petsc_interpolate() / interpolate_vars_on_mesh() already calls mesh.update_lvec() on both cache-hit and cache-miss paths (and empty coords no longer early-returns). This adds an extra collective globalToLocal() per evaluation and can be a noticeable performance hit in time-stepping loops. Consider removing this call here (or adding a mechanism to skip the internal update_lvec() when the caller has already performed it), relying on the now-parallel-safe petsc_interpolate() for the required collective participation.

Suggested change

	# CRITICAL: update_lvec() calls dm.globalToLocal() which is COLLECTIVE.
	# It MUST be called by ALL ranks before any rank enters petsc_interpolate,
	# because ranks with zero interior points would skip petsc_interpolate
	# (and its internal update_lvec call), deadlocking the ranks that do enter.
	mesh.update_lvec()
	# petsc_interpolate() is responsible for any required update_lvec()
	# / collective participation, including empty-coordinate cases.