pathfinding: LOW-severity cleanup backlog from the 2026-07-08 sweep round

[brendancol]

Is your feature request related to a problem?

The 2026-07-08 sweep round over xrspatial/pathfinding.py (issues #3644, #3645, #3649, #3651, #3652, #3655, #3657, #3660) fixed everything MEDIUM and above but left a trail of LOW findings in the sweep notes. None of them justifies a PR on its own, so this bundles them into one cleanup pass. Every item below was re-verified against current main on 2026-08-16. Items already fixed since the sweeps (TSP solver unit tests, anisotropic-cellsize golden tests) are not listed.

Performance nits (documented in the #3660 sweep, not fixed)

• _nearest_neighbor_2opt calls _tour_cost twice per 2-opt candidate, recomputing the full tour sum each time instead of the O(1) delta for a segment reversal. Only matters for hundreds of waypoints (_MAX_WAYPOINTS is 1000).
• _a_star_dask re-fetches the popped node's friction value (f_u_val, invariant across the loop) inside the 8-neighbor loop (pathfinding.py:752). Hoist above the loop.
• parent_ys/parent_xs are built with np.ones((h, w), dtype=np.int64) * NONE (pathfinding.py:286-287); np.full avoids the multiply pass.
• The dask start/goal crossability check runs two separate scalar .compute() calls (pathfinding.py:1069,1072); one dask.compute(start_val, goal_val) halves the scheduler round-trips.
• f_min for dask friction is recomputed with a full nanmin scan per a_star_search call (pathfinding.py:1130), so an N-waypoint multi_stop_search pays N-1 redundant scans. Flagged as a follow-up in multi_stop_search materializes the full grid on dask backends, defeating the sparse A* design #3660.

Error message (documented in the #3649 sweep, not fixed)

• A (0, 0)-size raster fails with ValueError: numpy.nanmin raises on a.size==0 and axis=None; So Bottleneck too. out of the resolution/f_min machinery. A shape check in _validate_surface_dims with a message naming the function and the empty input would match the other validators added in pathfinding: string barriers silently ignored, search_radius unvalidated, inconsistent start/goal and dims errors #3649. Reproduced on current main.

Test gaps (the "documented but not fixed" list from #3657)

• _ChunkCache eviction: no test exercises the maxsize overflow path, so the LRU popitem branch has zero coverage.
• _held_karp n==2 shortcut (pathfinding.py:1300-1301): the TSP unit tests added since use 3-4 cities only.
• The 2-opt improvement branch: test_nearest_neighbor_2opt_vs_held_karp's distance matrix never produces an improving swap (nearest-neighbor already returns the 2-opt-stable tour), so improved = True is never hit.
• _get_pixel_id default xdim/ydim arguments (pathfinding.py:103-106): every caller in the test suite passes explicit dims.

Benchmark gaps (documented in the #3645 sweep, not fixed)

• The asv pathfinding benchmarks route an open grid with no barriers and no friction, which is A*'s best case; regressions in barrier/friction handling or frontier growth won't show. Add a maze or barrier-field variant.
• dask+cupy is not a parameterizable backend anywhere in the suite (benchmarks/benchmarks/common.py get_xr_dataarray has no such type). This is suite-wide rather than pathfinding-specific; fixing it here would benefit every benchmark, but it can also be split out if it grows.

Describe the solution you'd like

One cleanup PR (or one PR plus a split-out for the suite-wide dask+cupy benchmark type):

• the five small performance fixes, none of which change results
• a zero-size raster validation error
• four targeted tests for the uncovered branches
• a barrier/friction benchmark variant

Excluded on purpose: the snap_start/snap_goal vs snap naming split (intentional per #3644, per-endpoint vs per-waypoint semantics) and the library-wide surface/raster/agg first-parameter naming (cross-cutting, tracked in #3644's notes).

Found by /sweeps-report rollup of the 2026-07-08 sweep round; re-verified 2026-08-16.