geotiff: keep no-georef int coords through to_geotiff round-trip (#1949)

[brendancol]

Closes #1949.

Summary

• _coords_to_transform now returns None when either spatial coord array has an integer dtype, so the read-side no-georef contract (open_geotiff windowed reads of non-georef TIFFs produce float64 half-pixel-shifted coords #1710, geotiff: GPU stripped windowed read of non-georef TIFF emits float64 coords (regression of #1710) #1753) survives a round-trip through to_geotiff.
• Before the fix, integer pixel coords [0, 1, 2, ...] produced a unit GeoTransform(origin_x=-0.5, origin_y=-0.5, pixel_width=1.0, pixel_height=1.0) that the writer emitted as ModelPixelScale / ModelTiepoint tags. On the next read the file looked georeferenced, the coord dtype silently flipped to float64, and a synthetic attrs['transform'] appeared.
• An explicit attrs['transform'] still wins because the writers consult it first; the helper only feeds the implicit-from-coords fallback.

Test plan

• New unit tests on _coords_to_transform for int64 / int32 / float / 3D-yxband.
• New round-trip tests through to_geotiff (eager numpy, dask streaming, GPU).
• Double-round-trip stability test.
• Explicit attrs['transform'] override test.
• Existing test_coords_to_transform_3d_1643.py, test_metadata_round_trip_1484.py, test_attrs_parity_1548.py, test_coord_regularity_1720.py, test_coords_1813.py, test_georef_edges.py still pass.
• Full geotiff test suite passes (9 pre-existing failures on main are unrelated to this change: test_predictor2_big_endian_gpu_1517.py, test_reader_kwarg_order_1935.py, test_size_param_validation_gpu_vrt_1776.py::TestWriteGeotiffGpuTileSize::test_tile_size_positive_works).

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[brendancol]

PR Review: geotiff: keep no-georef int coords through to_geotiff round-trip (#1949)

Blockers (must fix before merge)

• None

Suggestions (should fix, not blocking)

• xrspatial/geotiff/_coords.py:293 short-circuits before _is_regular(x, ...) runs. Integer arange coords are uniform so the regularity check would pass anyway, but if a caller ever passes hand-built int coords with non-uniform spacing (e.g. [0, 1, 3, 7]), the early return now hides that error rather than raising. A docstring note on coords_to_transform clarifying that integer dtype is treated as a no-georef sentinel (not just "uniform int coords") would head off confusion if someone later constructs int-indexed georeffed arrays.
• xrspatial/geotiff/tests/test_no_georef_writer_round_trip_1949.py:96 iterates (np.int32, np.int16, np.uint32) inside one test function. Splitting into pytest.mark.parametrize would surface which dtype fails when one regresses, instead of stopping at the first.

Nits (optional improvements)

• xrspatial/geotiff/tests/test_no_georef_writer_round_trip_1949.py:128 builds a 3D (y, x, band) array with int y/x coords but does not call to_geotiff on it. Extending the helper test into a round-trip on the 3D layout would close the loop with the existing 3D coverage in test_coords_to_transform_3d_1643.py.
• xrspatial/geotiff/tests/test_no_georef_writer_round_trip_1949.py:172 uses random.default_rng(seed=1949) for round-trip pixel data but never asserts pixel-value equality after round-trip. Adding np.testing.assert_array_equal(da.data, da2.data) would catch a regression where the writer truncates or rescales pixel bytes alongside the coord fix.
• xrspatial/geotiff/_coords.py:280 comment block is long enough that placing it inside a small helper docstring (or moving the rationale to the function docstring) would keep the function body easier to scan.

What looks good

• The fix lives in one place (_coords.py:293) and every writer path (eager, dask, GPU, VRT) flows through it, so the no-georef contract holds across backends without duplicate guards.
• kind in ('i', 'u') covers signed and unsigned integers at any width, matching the read-side np.int64 arange and any user-built int8/int16/uint32 variant.
• The explicit attrs['transform'] escape hatch is exercised by test_explicit_transform_attr_still_writes, confirming the helper only intercepts the implicit-from-coords fallback.
• Double round-trip stability test pins the fixed-point behaviour the issue was really after.
• GPU test is gated on cupy + CUDA availability and reuses the same assertions.

Checklist

• Algorithm matches reference/paper
• All implemented backends produce consistent results
• NaN handling is correct
• Edge cases are covered by tests
• Dask chunk boundaries handled correctly
• No premature materialization or unnecessary copies
• Benchmark exists or is not needed (not needed; metadata-only change)
• README feature matrix updated (if applicable) (N/A)
• Docstrings present and accurate

[brendancol]

PR Review: geotiff: keep no-georef int coords through to_geotiff round-trip (#1949) (re-review after rebase)

Rebased onto origin/main at 9ce0e60 (Windows writer-returns fix). Conflict-free. Pushed c996afc which converts the new test module's unconditional import tifffile into pytest.importorskip("tifffile") so collection no longer aborts on CI shards that don't ship tifffile, matching the pattern in test_miniswhite_nodata_1809.py and test_predictor3_big_endian.py. Two stray unused stdlib imports (os, tempfile) were dropped at the same time. Local pytest xrspatial/geotiff/tests/test_no_georef_writer_round_trip_1949.py reports 13 passed.

Blockers

• None

Suggestions

• None

Nits

• None

What looks good

• The integer-dtype short-circuit in coords_to_transform is the right layer for this fix: it sits before the uniformity check and leaves the explicit attrs['transform'] escape hatch intact, which the new test_explicit_transform_attr_still_writes pins.
• Coverage spans 2D, 3D, the dask streaming writer, a double round trip, and a GPU path gated on cupy availability, so all four backends are exercised.
• The expanded docstring on coords_to_transform calls out the no-georef sentinel and points to issues open_geotiff windowed reads of non-georef TIFFs produce float64 half-pixel-shifted coords #1710, geotiff: GPU stripped windowed read of non-georef TIFF emits float64 coords (regression of #1710) #1753, and geotiff: to_geotiff turns no-georef int64 pixel coords into a fake transform on round-trip #1949, which makes the rationale discoverable from the helper.
• dtype.kind in ('i', 'u') covers signed and unsigned widths uniformly; the parametrised test_coords_to_transform_returns_none_for_int32_coords locks that contract in.

Checklist

• Rebase clean on origin/main
• Targeted test green locally (13/13)
• CI collection regression on ubuntu-3.12 / macos / windows addressed via importorskip
• Force-with-lease push to origin
• CI re-run on the new head

[brendancol]

Rebased onto current main at 8da3ca7 to pick up the Python 3.14 CI matrix limit (b1579f8).