geotiff: to_geotiff silently strips georef on int64 step-1 user coords

[brendancol]

Describe the bug

to_geotiff silently strips georef from any DataArray whose x and y coords are contiguous int64 step-1 arrays. _is_no_georef_sentinel (xrspatial/geotiff/_coords.py:39-77) accepts every ascending step-1 int64 coord array as the read-side no-georef placeholder, but real user grids match that exact shape.

Reproducer

import numpy as np, xarray as xr
from xrspatial.geotiff import to_geotiff, open_geotiff

x = np.array([500, 501, 502], dtype=np.int64)
y = np.array([1000, 1001], dtype=np.int64)
da = xr.DataArray(np.zeros((2, 3), dtype=np.float32),
                  coords={'y': y, 'x': x}, dims=('y', 'x'),
                  attrs={'crs': 4326})

to_geotiff(da, '/tmp/repro.tif')
out = open_geotiff('/tmp/repro.tif')
print(out.x.values)  # -> [0, 1, 2]  georef silently lost

Expected behaviour

The output should round-trip with the original coords, or to_geotiff should raise so the user can supply attrs['transform'] explicitly. Silent metadata loss in a foundational IO path is worse than a noisy error.

Root cause

_is_no_georef_sentinel was tightened in #2087 from "any integer dtype" to "exact arange pattern" to fix an earlier silent-strip bug. The new check is still purely shape-based. The docstring at lines 63-68 admits the trade-off: a user-authored int64 step-1 grid still matches.

Fix sketch

Require an explicit no-georef marker instead of inferring from coord shape. Two options:

• (a) attrs['_xrspatial_no_georef'] = True. The reader stamps it when emitting placeholder coords; the writer checks it before treating coords as no-georef.
• (b) Stamp a private flag on the coord variable (coord.attrs) that the reader sets and the writer checks. Survives ops that preserve coord attrs.

With either fix, int64 coords without the marker should fall through to the normal coords_to_transform path: synthesise a real transform, or raise NonUniformCoordsError on irregular spacing.

Environment

• xrspatial main (commit 22aa43c)
• python 3.11, numpy 2.x