geotiff: to_geotiff turns no-georef int64 pixel coords into a fake transform on round-trip

[brendancol]

A round-trip of a no-georef TIFF through open_geotiff -> to_geotiff -> open_geotiff silently injects a synthetic attrs['transform'] and switches the y/x coord dtype from int64 to float64.

The read side documents the no-georef contract: integer pixel coords [0, 1, 2, ...] with int64 dtype, no attrs['transform'] (#1710, #1753). The write side does not honour the same signal -- _coords_to_transform infers a unit GeoTransform(origin_x=-0.5, origin_y=-0.5, pixel_width=1.0, pixel_height=1.0) from [0, 1, 2, ...] coords and writes it as real ModelPixelScale / ModelTiepoint tags. On the next read the file looks like a georeferenced raster and the no-georef branch never fires.

Reproduction

import numpy as np
import tempfile, os
import tifffile
from xrspatial.geotiff import open_geotiff, to_geotiff

with tempfile.TemporaryDirectory() as td:
    src = os.path.join(td, "src.tif")
    arr = np.arange(20, dtype=np.float32).reshape(4, 5)
    tifffile.imwrite(src, arr, photometric='minisblack', planarconfig='contig')

    da = open_geotiff(src)
    print(da.coords['y'].dtype, da.coords['y'].values)
    # int64 [0 1 2 3]
    print('transform' in da.attrs)
    # False

    out = os.path.join(td, "out.tif")
    to_geotiff(da, out, compression='none')

    da2 = open_geotiff(out)
    print(da2.coords['y'].dtype, da2.coords['y'].values)
    # float64 [0. 1. 2. 3.]
    print(da2.attrs.get('transform'))
    # (1.0, 0.0, -0.5, 0.0, 1.0, -0.5)

Impact

• Cat 2 (coords): coord dtype silently changes between input and output (HIGH per sweep criteria).
• Cat 1 (attrs): writer adds an inferred transform attr the input never had (MEDIUM).
• Downstream code that branches on 'transform' in da.attrs to detect georef status will mis-classify the round-tripped file as georeferenced.

Cause

xrspatial/geotiff/_writers/eager.py:519, _writers/eager.py:839, _writers/gpu.py:350 all fall back to _coords_to_transform(data) when attrs.get('transform') is missing. _coords_to_transform (_coords.py:235) reads da.coords[xdim].values, computes pixel_width = float(x[1] - x[0]), and returns GeoTransform(...) -- with no check for the integer-dtype signal that the read side uses to mean "no georef".

Affects every writer path: to_geotiff eager, to_geotiff dask streaming (via the same fallback), to_geotiff(vrt=...) per-tile, write_geotiff_gpu. The GPU dask+cupy writer goes through the same helper.

Proposed fix

_coords_to_transform returns None when either spatial coord array has an integer dtype (which only happens from the read-side no-georef fallback). Float coord arrays continue to produce a GeoTransform. Callers already handle None correctly: write() skips the geokey emission when geo_transform is None.