resample: reject irregular/non-monotonic spatial coordinates

[brendancol]

Closes #2663

Problem

resample assumes a regular, monotonic raster but never checks for one. target_resolution derives the input resolution from calc_res() (full-extent (max - min) / (n - 1)), while the output coordinates are rebuilt from first/last neighbour spacing in _new_coords. On an irregular grid those two views of resolution disagree, so the output width and coordinate range stop matching the input. With x=[0, 1, 4] and target_resolution=1.0 the output had width 6 and x coords [0, 1, 2, 3, 4, 5], spilling past the input range [0, 4].

Change

• Add _validate_monotonic_regular_coords, called at the top of resample(), which rejects spatial coordinates that are non-monotonic or unevenly spaced with a clear ValueError.
• Strictly-monotonic check accepts both ascending and descending (north-up) axes. Even-spacing check uses a relative tolerance so floating-point jitter passes.
• Inputs without spatial coords, or with a length-1 axis, are left alone (nothing to validate).

Backends

Coordinate validation reads numpy-backed coords regardless of the data backend, so the guard behaves identically across numpy, cupy, dask+numpy, and dask+cupy. Tests parametrize over every available backend.

Test plan

• Irregular and non-monotonic x/y rejected (all available backends)
• Regular ascending, descending-y, and float-jitter grids still work
• Missing-coords and single-pixel-axis inputs unaffected
• Existing resample suite still green (321 passed)

[brendancol]

PR Review: resample: reject irregular/non-monotonic spatial coordinates

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

• The even-spacing tolerance combines rtol=1e-5 with atol=abs(step) * 1e-5 (xrspatial/resample.py:236), so both terms scale with step. It works, but the atol term is redundant: np.allclose already uses atol + rtol * |b| with b = step, so rtol alone scales by |step|. Either drop the atol term or keep it as a deliberate small floor. Behavior is correct as written.

Nits (optional improvements)

• _validate_monotonic_regular_coords re-runs the coord checks once per band on 3D inputs, because resample() recurses per-band and revalidates. Harmless (the coords are identical and the check is cheap), but a one-line note would make the redundancy explicit.

What looks good

• The guard runs before any geometry math or backend dispatch, so the bad-geometry path is closed at the source.
• A descending y-axis (north-up rasters) is correctly accepted via all(diffs > 0) or all(diffs < 0).
• Missing-coords and length-1 axes are skipped rather than erroring, matching the existing fall-through behavior.
• Coords are numpy-backed in xarray regardless of data backend, so one check covers all four backends. The tests parametrize over every available backend anyway.
• Error messages name the offending dimension and state the contract.

Checklist

• Behavior matches the documented contract (regular monotonic rasters only)
• All implemented backends produce consistent results (coord check is backend-independent)
• NaN handling unaffected (validation is on coords, not data)
• Edge cases covered by tests (missing coords, single-pixel axis, float jitter, descending y)
• Dask chunk boundaries: not applicable, no chunked compute added
• No premature materialization (coords are already eager numpy)
• Benchmark: not needed, validation-only change
• README feature matrix: not applicable, no new function or backend change
• Docstring updated (Raises section)

[brendancol]

Follow-up review (after commit 2df9efc)

Both items from the first pass are resolved:

• Spacing tolerance: the atol=abs(step) * 1e-5 term is gone. The check now uses np.allclose(diffs, step, rtol=1e-5, atol=0.0), so a single rtol controls the tolerance and it tracks the coordinate magnitude. The float-jitter test still passes.
• Per-band re-validation: documented in the helper docstring as a cheap, harmless repeat on identical coords.

No new findings. Tests: 240 passed across the resample suite (numpy + dask backends; cupy backends gated by availability).

Checklist

• Behavior matches the documented contract
• Backend-independent coord check, parametrized tests
• NaN handling unaffected
• Edge cases covered
• No premature materialization
• Docstring updated

[brendancol]

CI fix (commit 46bbcc9)

The first CI run failed: the new test file gated its cupy backends with cuda_and_cupy_available() / dask_array_available() imported from general_checks, but those names are pytest.mark.skipif MarkDecorators, not boolean predicates. Calling a MarkDecorator is always truthy, so the cupy and dask+cupy backends were added unconditionally and those parametrized cases ran on CI (where cupy is not installed), failing at import cupy. It passed locally only because cupy is present here, so the cases ran on real hardware and masked the bug.

Fix: switch to the boolean predicates has_cuda_and_cupy() and has_dask_array() from xrspatial.utils (the convention used in test_reproject.py). With cupy absent, BACKENDS is now ['numpy', 'dask+numpy'] and no cupy cases are collected. Verified by simulating cupy-absent locally.

Post-fix CI: pytest (ubuntu/windows/macos 3.14), pytest-cog-validator (3.12), pytest-geotiff-corpus, and Label PRs all green. Resample suite: 364 passed locally.