Fix COG cubic overview poisoned by nodata sentinel (#1623)

[brendancol]

Summary

Closes #1623.

to_geotiff(cog=True, overview_resampling='cubic', nodata=<finite>) on a float raster with NaN regions produced overview pixels with bad ringing near nodata borders. Same root cause as #1613, but for the cubic branch: the writer rewrote NaN to the sentinel before reduction, then _block_reduce_2d(method='cubic') ignored nodata and handed the sentinel-poisoned array to scipy.ndimage.zoom(order=3). The cubic spline blended the sentinel into neighbouring cells (values like 1133.44 and -10290.08 appeared where the source data was a constant 100).

Fix:

• CPU _block_reduce_2d cubic branch: when nodata is supplied on a float dtype and the sentinel is found, mask sentinel to NaN, run cubic with prefilter=False so a single NaN cannot poison the whole row/column, then rewrite any NaN in the result back to the sentinel. prefilter=False only fires when a sentinel is present, so the non-nodata cubic semantics are unchanged.
• GPU _block_reduce_2d_gpu: previously raised ValueError on method='cubic'. Added a CPU fallback (same pattern as 'mode') so the GPU writer produces byte-equivalent overviews to the CPU writer. GPU_OVERVIEW_METHODS now includes 'cubic'.

12 regression tests in test_cog_cubic_overview_nodata_1623.py cover the helper, the round trip, the no-nodata regression guard, +/-inf nodata, the NaN-sentinel no-op, the GPU fallback, and CPU/GPU byte-equivalence.

Test plan

• pytest xrspatial/geotiff/tests/test_cog_cubic_overview_nodata_1623.py (12 pass, including 3 GPU-gated)
• pytest xrspatial/geotiff/tests/test_cog_overview_nodata_1613.py (14 pass, the prior nodata-overview fix)
• pytest xrspatial/geotiff/tests/test_cog.py xrspatial/geotiff/tests/test_overview_filter.py xrspatial/geotiff/tests/test_streaming_codecs_2026_05_11.py (44 pass, existing cubic round-trip + cubic-distinct-from-mean)
• Full xrspatial/geotiff/tests/ run: 1256 pass, 7 skipped, 3 pre-existing matplotlib-deepcopy failures in TestPalette (unrelated)

State file .claude/sweep-accuracy-state.csv updated with the pass-16 record for the geotiff module.

[Copilot]

Pull request overview

This pull request fixes incorrect COG overview pixels produced when using cubic resampling with a finite nodata sentinel on float rasters that contain NaN regions (issue #1623). It ensures both CPU and GPU writer paths avoid “sentinel ringing” artifacts near nodata borders and remain consistent with each other.

Changes:

• Update CPU _block_reduce_2d(..., method='cubic') to honor nodata for float arrays by masking sentinel→NaN, running cubic with prefilter=False, then rewriting NaNs back to the sentinel.
• Add a GPU cubic overview implementation via CPU fallback and include 'cubic' in GPU_OVERVIEW_METHODS.
• Add a dedicated regression test module covering CPU helper behavior, end-to-end COG round trips, edge cases (+/-inf, NaN sentinel), and CPU/GPU equivalence.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
xrspatial/geotiff/_writer.py	Implements nodata-aware cubic downsampling to prevent sentinel contamination during SciPy cubic zoom.
xrspatial/geotiff/_gpu_decode.py	Adds 'cubic' to GPU overview methods and routes cubic overviews through a CPU fallback for parity.
xrspatial/geotiff/tests/test_cog_cubic_overview_nodata_1623.py	Adds regression coverage for cubic+nodatavalue behavior and CPU/GPU equivalence.
.claude/sweep-accuracy-state.csv	Updates sweep/inspection state record for the geotiff module.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[brendancol]

@copilot apply changes based on this feedback

[Copilot]

Updated in 7b81662: the _block_reduce_2d docstring now scopes the NaN-preservation wording to the nan-aware reducers and explicitly notes that the cubic path rewrites NaNs back to the sentinel before returning.