test(rasterize): close eager-cupy merge-mode coverage gap (deep-sweep pass 3)

[brendancol]

Summary

Deep-sweep test-coverage pass 3 on xrspatial.rasterize. Adds 23 tests in test_rasterize_coverage_2026_05_27.py, all passing on a CUDA host. No source changes.

• Cat 1 HIGH -- eager cupy (use_cuda=True, no chunks=) had no parametrised six-mode merge parity test against eager numpy. TestDaskNumpy.test_merge_mode_parity and TestDaskCupy.test_merge_mode_parity carry it for the tiled paths, and pass 1 only pinned merge='last' on a single non-overlapping polygon. A routing regression in any of the six GPU atomic kernels (_ensure_gpu_kernels, rasterize.py:1308-1556) would slip past the dask+cupy parity tests because the tiled finalize path is different. Pin polygon and point overlap scenes across last / first / max / min / sum / count, with sanity checks (first != last, min < max) that lock the fixture as non-degenerate.
• Cat 1 MEDIUM -- _run_cupy with zero geometries had no direct test (eager numpy and both dask backends did). Pin that use_cuda=True with [] keeps the cupy backend instead of short-circuiting to numpy, under both an explicit fill and the default NaN fill.
• Cat 2 MEDIUM -- merge='count' with all-equal property values: four overlapping rectangles burning the same 1.0 must still count overlaps as > 1 on all four backends. A future GPU atomic optimisation that deduplicated identical-value writes would silently divide counts by the number of unique values.
• Cat 4 MEDIUM -- name= kwarg thread-through on dask+numpy / eager cupy / dask+cupy (eager numpy was the only path covered).

State CSV updated: rasterize row -> last_inspected=2026-05-27, severity_max=HIGH, categories_found=1;2;4.

Test plan

• pytest xrspatial/tests/test_rasterize_coverage_2026_05_27.py -- 23 passed locally on CUDA host
• pytest xrspatial/tests/test_rasterize*.py -- 325 passed, 2 skipped (no regressions in passes 1+2 or the main test file)
• CI green on numpy-only and dask-only runners (no GPU markers must be hit there)

[brendancol]

One blocker, two nits. Fix in 99b9b3e.

Blocker -- module-level box(...) crashes import without shapely

_EAGER_CUPY_MERGE_PAIRS, _ALL_EQUAL_PAIRS, and TestNameKwargBackends._SIMPLE_PAIR all called box(...) at module scope. Since #2496 made shapely optional, a numpy-only runner hits NameError: name 'box' is not defined during collection, before pytestmark skip has a chance to fire.

Reproduced by injecting ImportError on import shapely.* and running pytest --collect-only: prior tip failed, HEAD collects all 23 tests cleanly.

Fix: wrap the two list constants in if has_shapely: and turn the class attribute into a _simple_pair() method so box is only touched from test bodies.

Nits (not pushed)

1. test_eager_cupy_first_differs_from_last and test_eager_cupy_min_differs_from_max sit under @skip_no_cuda but only call the numpy backend. The non-degeneracy property they pin doesn't depend on the backend, so they'd still be useful on numpy-only runners. Not worth re-shuffling.

2. test_count_numpy asserts max >= 3.0; observed max is 4.0 (four-way overlap at world (5.5, 5.5)). == 4.0 would catch a regression that dropped one of the four writes, but the >= 3.0 bound is what the test was actually built to defend against (dedup-to-1), so leaving it.

Test status

• new file: 23 passed (CUDA host)
• new file + test_rasterize.py: 238 passed, 2 skipped
• --collect-only with shapely suppressed: 23 collected, no NameError