rasterize: propagate NaN burns under max/min on every backend (#2255)

[brendancol]

Closes #2255.

Summary

• GPU paths initialised the max/min output buffer to -inf / +inf and merged with atomicMax / atomicMin. NaN comparisons in CUDA return False, so NaN burns never displaced the inf initialiser, and any finite competitor silently dropped the NaN.
• CPU paths had the same bug in milder form: a NaN burn that arrived after a finite burn was silently dropped because 'NaN > finite' is False.
• All four backends now propagate NaN strictly: any NaN burn, regardless of input order, poisons the pixel under max / min. CPU checks the burn for NaN explicitly. GPU repurposes the 'written' buffer (already threaded through every kernel) as an H*W int32 NaN-mask; atomicMax(mask, 1) records NaN burns and _gpu_finalize_buffers stamps NaN into matching pixels.

Backend coverage

• numpy, cupy, dask+numpy, dask+cupy all share the new contract.

Test plan

• pytest xrspatial/tests/test_rasterize_coverage_2026_05_21.py (81 passed, includes the new TestNaNPropagationMaxMin class — 28 parametrised tests across the four backends).
• pytest on the full rasterize suite: test_rasterize.py, test_rasterize_accuracy.py, test_rasterize_gpu_race_2167.py, test_rasterize_all_touched_supercover_2169.py, test_rasterize_tile_props_slice_2020.py, test_rasterize_signature_annot_2250.py, test_rasterize_coverage_2026_05_17.py — 371 passed, 2 skipped.
• Live cupy / dask+cupy reproducer from the issue now returns NaN on both backends.

[brendancol]

PR Review: rasterize: propagate NaN burns under max/min on every backend (#2255)

Blockers

None.

Suggestions

None.

Nits

• xrspatial/rasterize.py:102-106 — the if pixel != pixel: return pixel branch in _merge_max is dead code. The earlier val != val early-return rules out NaN burns reaching this point, and the subsequent val > pixel already returns pixel when pixel is NaN (because finite > NaN is False). The comment says it is kept "to document intent and keep the contract symmetric with the GPU path", which is fair, but a single inline comment on the val > pixel check would do the same job without the dead branch. Same in _merge_min at line 119-120.
• xrspatial/tests/test_rasterize_coverage_2026_05_21.py — TestNaNPropagationMaxMin is appended to a file whose module docstring claims "The 'fix' in this sweep is adding tests. No source changes." That claim is no longer accurate with this PR's source-side fix. Either update the docstring or move the new class to its own file (e.g. test_rasterize_nan_propagation_2255.py).

What looks good

• Small diff: 74/-20 in rasterize.py, 84/-65 in the test file. No drive-by refactoring.
• Repurposing written as the NaN mask for max/min avoids changing any kernel signatures — the buffer already flows through every kernel.
• The int8 → int32 dtype bump is documented in _gpu_init_buffers with the reason (cuda.atomic.max does not support int8).
• New tests are parametrised across all four backends via ALL_BACKENDS.
• Cases covered: NaN-first overlap, finite-then-NaN (the milder CPU-side bug), single NaN burn, min variant, untouched-pixel fill preservation.

Checklist

• Algorithm matches the issue's stated contract — strict NaN propagation on max/min, regardless of input order, on all four backends.
• Four backends produce consistent results — seven new tests parametrised across numpy / cupy / dask+numpy / dask+cupy, all pass.
• NaN handling verified against the issue's reproducer on cupy and dask+cupy. test_finite_then_nan_max[numpy] would have failed on the old CPU kernel.
• Edge cases covered: untouched pixels, partial NaN/finite coverage, min variant, single NaN burn.
• Dask handled correctly — tiles partition the output, so per-tile _gpu_init_buffers / _apply_merge_gpu / _gpu_finalize_buffers is sufficient.
• No premature materialization. cupy.where in finalize allocates once for the NaN stamp and once for the fill restore.
• Benchmark not needed — no new public function, and the inner-loop cost is one extra val == val check.
• README feature matrix not applicable — behaviour change, no API addition.
• Docstrings updated for _gpu_init_buffers and _gpu_finalize_buffers.

[brendancol]

Follow-up review pass

Both nits from the prior pass are addressed in 54fb381:

• Dead NaN branch in _merge_max / _merge_min — fixed. Dropped the explicit if pixel != pixel: return pixel check and left a short comment on the val > pixel line documenting that a NaN pixel from an earlier burn is still preserved (because val > NaN is False, the function falls through to return pixel).
• New tests in the deep-sweep coverage file — fixed. Moved TestNaNPropagationMaxMin out of test_rasterize_coverage_2026_05_21.py into a dedicated test_rasterize_nan_propagation_2255.py. The deep-sweep file's docstring ("No source changes") is no longer mis-applied.

Verification

• pytest xrspatial/tests/test_rasterize_nan_propagation_2255.py xrspatial/tests/test_rasterize_coverage_2026_05_21.py -q — 81 passed (28 NaN propagation tests in the new file + 53 in the coverage file).
• pytest xrspatial/tests/test_rasterize.py -q — 215 passed, 2 skipped. No regressions.

Disposition summary

Original finding	Disposition
Blockers	None to begin with
Suggestions	None to begin with
Nit 1: dead NaN branch	Fixed in 54fb381
Nit 2: tests in wrong file	Fixed in 54fb381

No remaining findings.