What the tests check

The tests use GDAL as an oracle. Small synthetic files and GDAL’s output for them live in wxylib/fixtures, made by scripts there with GDAL and zarr-python.

• Reads (wxy-tiff, wxy-zarr, wxy-virtual): every window must match what GDAL reads, byte for byte, across five GeoTIFFs (int16, uint8, uint16, float32; DEFLATE, ZSTD, LZW with predictor; COG and plain; pixel and band interleaved; BigTIFF), three of them converted to Zarr, and a byte-reference table over three netCDF4 files (one chunk missing on purpose, which must read as fill).
• Warps (wxy-warp, wxy-cli): three source/target pairs (global to Antarctic polar stereographic, polar to lon/lat, UTM to lon/lat) with nearest, bilinear and cubic, at 1:1 and coarser, compared with what GDAL’s warper writes; curvilinear swaths placed by cell centres or corners, against GDAL’s geolocation-array warp; and an OISST-style 4D store as xarray writes it, day by day.
• Extraction (wxy-warp): values at random points, by cell and interpolated, against GDAL’s point reads; and the cells under random polygons with holes, by cell centre and all-touched, against GDAL’s rasterizer.
• Model (wxy-domain, wxy-geozarr): unit tests, including real-world grid cases from regular-in-what (a float32 GHRSST axis, AVISO’s Mercator latitude, NSIDC polar stereographic).

The bar is identity: every warp, point read and polygon case must give GDAL’s numbers exactly, cell for cell.

CI runs cargo test --workspace against the fixtures on every push and pull request to core.