Add golden-value reference tests for a_star_search (#3655)

[brendancol]

Closes #3655

• Adds golden-value regression tests for a_star_search: goal costs on a 6x6 open grid (8- and 4-connectivity, matches the closed form), an 8x8 barrier maze (including a no-path pair), a 6x6 friction surface, and an anisotropic-cellsize grid.
• Expected values were generated on 2026-07-08 and verified to machine precision against scipy 1.16.1 sparse.csgraph.dijkstra and scikit-image 0.26.0 MCP_Geometric (same edge model: geometric distance, or distance times mean endpoint friction). The values are hardcoded, so neither reference library is needed at test time.
• Test-only; no source changes. The full comparison harness (64x64 inputs, NaN friction holes, path-chain consistency, all four backends) is in Add golden-value reference tests for a_star_search (validated against scipy dijkstra and skimage MCP_Geometric) #3655.

Backend coverage: the new tests run on numpy and dask+numpy. The existing test_cupy_matches_numpy / test_dask_cupy_matches_numpy parity tests carry the pinned values to the GPU backends.

Also updates the reference-validation sweep state CSV row for pathfinding (verdict MATCHES).

Test plan:

• pytest xrspatial/tests/test_pathfinding.py -k golden (18 passed)
• pytest xrspatial/tests/test_pathfinding.py (75 passed, includes cupy and dask+cupy on a CUDA host)
• flake8 on the test file: same 14 pre-existing findings before and after

https://claude.ai/code/session_0155N4QGamQVxgpAAPbpQNq4

[brendancol]

PR Review: Add golden-value reference tests for a_star_search (#3655)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

• xrspatial/tests/test_pathfinding.py:1193: _golden_raster overlaps with the existing _make_raster helper a few hundred lines up. The overlap is deliberate and correct: _make_raster writes attrs={'res': res} with res[0] used as the y spacing, while get_dataarray_resolution (utils.py:685) reads attrs['res'] as (cellsize_x, cellsize_y). With square cells that mismatch is invisible; with the anisotropic case in this PR it would flip cx and cy. _golden_raster avoids the trap by setting no res attr, so resolution comes from the coords. A one-line comment in _golden_raster saying why it does not reuse _make_raster would keep someone from "simplifying" this later and silently breaking the anisotropic test.

Nits (optional improvements)

• xrspatial/tests/test_pathfinding.py:1203-1226: the case tuple element exists only to name the parametrization. pytest.param(..., id=...) expresses that directly and drops the unused function argument. Purely cosmetic; the current form is readable as-is.

What looks good

• Golden values are traceable: the block comment names both reference tools with exact versions (scipy 1.16.1 dijkstra, scikit-image 0.26.0 MCP_Geometric), the generation date, and the observed agreement (2e-15 rel or better), and the issue carries the full runnable comparison script.
• Goal-cost pinning is the right invariant: optimal cost is unique even when the optimal path is not, so the tests will not flake on heap tie-breaking changes.
• Maze and friction grids are written out as literals rather than regenerated from RandomState, so the tests do not depend on numpy RNG stream stability.
• Coverage picks the spots where a cost-model regression would hide: 4- vs 8-connectivity, barriers, friction weighting, anisotropic cellsize, and a no-path case asserting the all-NaN contract.
• Runs on numpy and dask+numpy; the existing cupy/dask+cupy parity tests extend the pinned values to GPU backends without duplicating them here.
• Test-only diff plus the sweep state CSV row; no source or dependency changes.

Checklist

• Algorithm matches reference/paper (verified against two independent implementations plus closed form)
• All implemented backends produce consistent results (numpy/dask direct; cupy via existing parity tests)
• NaN handling is correct (no-path case asserts all-NaN output)
• Edge cases are covered by tests (no-path, anisotropic cells, 4-connectivity)
• Dask chunk boundaries handled correctly (dask param uses non-trivial chunks)
• No premature materialization or unnecessary copies (test code; .values on small grids is fine)
• Benchmark exists or is not needed (test-only change; #3645 tracks pathfinding benchmarks separately)
• README feature matrix updated (not applicable, no new function)
• Docstrings present and accurate (not applicable: no public API change)

[brendancol]

PR Review: follow-up after review commits

Both findings from the first review are addressed in 7847ffe:

• Suggestion (fixed): _golden_raster now carries a comment explaining why it does not reuse _make_raster (the attrs['res'] ordering mismatch with get_dataarray_resolution that would flip cx/cy on anisotropic cells).
• Nit (fixed): the golden parametrizations use pytest.param(..., id=...); the unused case argument is gone.

One process note: an intermediate commit accidentally rewrote the test file's CRLF line endings to LF, which ballooned the diff to ~1350 changed lines. That was caught and amended; the follow-up commit is back to a 23+/18- diff with the file's original CRLF endings intact.

Re-verified after the changes: 75 passed in xrspatial/tests/test_pathfinding.py, flake8 finding count unchanged (14 pre-existing, none in the added block).

No remaining findings.