geotiff: cross-library accuracy test (rasterio vs xarray-spatial vs zarr round-trip) #1961

Test idea

Read the same GeoTIFF three ways and assert the results agree:

1. rasterio.open(path) → numpy array, transform, CRS, nodata.
2. xrspatial.geotiff.open_geotiff(path) → xr.DataArray.
3. Write the xarray-spatial DataArray with .to_zarr(...), reopen with xr.open_zarr(...).

Pairwise checks across the three readers:

• pixel values (array_equal for integer dtypes, allclose with a documented tolerance for floats)
• shape, dtype, nodata sentinel
• x and y coordinate arrays within float tolerance
• CRS WKT after normalising through _resolve_crs_to_wkt
• affine transform (origin, pixel size) within 1e-9 relative

Why

The project already has backend-parity tests inside its own implementation (numpy / cupy / dask / dask+cupy). What's missing is a test that pins the GeoTIFF reader against an external reference (rasterio) and against a round-trip through a different on-disk format (Zarr). A regression in header parsing, georef extraction, coord generation, nodata handling, or Zarr metadata propagation can pass every existing test and only surface when a user files a bug.

Suggested inputs

Cover the cases that have drifted before:

• single-band float32 with a non-NaN nodata sentinel
• multi-band uint16 with per-band nodata
• north-up and south-up rasters (negative vs positive pixel_height)
• 1×N / N×1 stripe (geotiff: 1xN / Nx1 georeferenced writes silently strip transform #1945)
• tiled COG, no overviews
• no-georef raster with integer coords (geotiff: to_geotiff turns no-georef int64 pixel coords into a fake transform on round-trip #1949)

Suggested location

xrspatial/geotiff/tests/test_round_trip_parity_rasterio_zarr_<issue#>.py

Gate the imports with pytest.importorskip("rasterio") and pytest.importorskip("zarr") so the CPU-only matrix without those deps stays green.

Out of scope

• performance comparison (handled by the bench skill)
• GPU read paths (covered by the existing backend parity tests)
• cloud / HTTP sources (local-file only)