QA/QC: validate sieve output against GDAL/rasterio reference

[brendancol]

Reason or problem

The sieve module (xrspatial/sieve.py) was added recently and claims equivalence with GDAL's gdal_sieve.py. The existing tests cover internal correctness (region sizes, connectivity, skip_values, NaN handling), but nothing actually compares output against GDAL's sieve. Without that cross-check, we don't know if the behavior matches what users coming from GDAL/rasterio expect.

Proposal

Add a QA/QC test module that runs xrspatial.sieve() and rasterio.features.sieve() on identical inputs and asserts matching output. Cover:

• Multiple thresholds (1, 5, 10, 50)
• Both connectivity modes (4 and 8)
• Rasters with NaN/nodata regions
• Integer vs float inputs
• A realistic classified raster, not just toy grids
• Edge cases: single-value raster, checkerboard pattern, L-shaped regions

Where outputs differ, document why (e.g., tie-breaking order, nodata semantics).

Design

• Test file: xrspatial/tests/test_sieve_gdal_parity.py
• Use rasterio.features.sieve as reference (already an optional dep in CI)
• Generate synthetic rasters in test fixtures so tests are self-contained
• Use pytest.importorskip("rasterio") so tests skip if rasterio isn't installed
• Document behavioral differences in inline comments, not a separate doc

Value

• Lets us verify that migrating from GDAL sieve to xrspatial sieve produces the same results
• Catches regressions if our output drifts from GDAL's behavior
• Documents intentional differences

Drawbacks

• Adds rasterio as a test-time dependency for this file (already available in CI)
• If GDAL changes sieve behavior in future versions, tests may need updating

Alternatives

• Compare against scipy.ndimage.label + manual merge, but that just reimplements the algorithm rather than validating against the accepted standard
• Use GDAL python bindings directly, but rasterio.features.sieve is the more common user-facing API

Unresolved questions

• Tie-breaking: GDAL may pick a different neighbor when multiple neighbors have the same size. We should document these cases rather than trying to match GDAL's arbitrary tie-breaking.