fix(dynamics): mask stale neighbor matrix entries before COO conversion

[laserkelvin]

Summary

• Fix stale edges in COO neighbor list output when neighbor counts shrink between Verlet skin rebuilds
• Cache _col_range tensor to avoid per-step allocation in _update_edges_group
• Add TestStaleCOOEntries regression test covering the shrinking-neighbors scenario

Problem

When NeighborListHook uses a Verlet skin for selective rebuilds, the upstream warp kernel (_update_neighbor_matrix) only writes slots 0..num_neighbors[i]-1 in the neighbor matrix. If an atom's neighbor count decreases between rebuilds, the old entries in neighbor_matrix[i, new_count:old_count] are never overwritten with fill_value — they retain stale atom indices from the previous build.

get_neighbor_list_from_neighbor_matrix uses neighbor_matrix != fill_value to identify valid entries, so these stale indices pass the mask and appear as phantom edges in the COO output. This causes energy/force mismatches during MD compared to ASE + model.

Fix

In _update_edges_group, before calling get_neighbor_list_from_neighbor_matrix, create a boolean mask col_range >= num_neighbors[i] and overwrite stale slots with fill_value in-place. The col_range tensor is cached in _alloc_output_tensors alongside the neighbor matrix buffers.

The in-place overwrite is safe because the next rebuild will rewrite all valid slots via atomic_add.

Test

TestStaleCOOEntries sets up a 3-atom line system, runs an initial rebuild, then moves an atom far enough to trigger a Verlet skin rebuild with fewer neighbors. It verifies:

1. No stale edges remain in the COO edge_index
2. Per-atom COO edge counts match the internal num_neighbors buffer

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[laserkelvin]

/ok to test d8824dd

[greptile-apps]

Greptile Summary

This PR fixes phantom COO edges caused by stale entries in the neighbor matrix when neighbor counts shrink between Verlet skin rebuilds. The fix in _update_edges_group masks all slots ≥ num_neighbors[i] with fill_value in-place before passing the matrix to get_neighbor_list_from_neighbor_matrix, and caches _col_range to avoid per-step allocation. A regression test covering the shrinking-neighbor scenario is included.

Important Files Changed

Filename	Overview
nvalchemi/hooks/neighbor_list.py	Core fix: in-place stale-entry masking added to _update_edges_group before COO conversion; _col_range properly allocated in _alloc_output_tensors and initialised to None in __init__.
test/dynamics/test_neighbor_list_hook.py	New TestStaleCOOEntries class correctly validates the shrinking-neighbor scenario and verifies COO edge counts against the internal _num_neighbors buffer.

_{Reviews (5): Last reviewed commit: "Merge remote-tracking branch 'origin/mai..." | Re-trigger Greptile}

[laserkelvin]

/ok to test 0f6d192

[laserkelvin]

/ok to test d767402

[laserkelvin]

/ok to test 352a160

[laserkelvin]

/ok to test 57883de