geotiff: extract_geo_info float-only nodata parse loses uint64 max / int64 max sentinels

[brendancol]

Summary

The reader's GDAL_NODATA parsing in _geotags.extract_geo_info (and the related helpers _reader._resolve_masked_fill / _reader._sparse_fill_value) parses the nodata string via float() unconditionally. For 64-bit integer rasters whose declared sentinel is uint64 max (2**64 - 1) or int64 max (2**63 - 1), the float64 round-trip rounds the sentinel to a value that exceeds the dtype's representable range. The downstream integer-mask gate info.min <= int(nodata) <= info.max then rejects the cast and the sentinel pixel survives as a literal valid integer rather than being masked to NaN.

Repro:

import numpy as np
import xarray as xr
import xrspatial.geotiff as g
import tempfile, os

with tempfile.TemporaryDirectory() as d:
    arr = np.full((16, 16), 100, dtype=np.uint64)
    arr[0, 0] = 2**64 - 1  # uint64 max as sentinel
    da = xr.DataArray(arr, dims=['y','x'])
    p = os.path.join(d, 't.tif')
    g.to_geotiff(da, p, nodata=2**64-1)
    out = g.open_geotiff(p)
    print(out.dtype, out.values[0, 0])  # uint64 18446744073709551615
    # expected: float64 NaN, after sentinel-to-NaN promotion

Same failure pattern for np.int64 with nodata = 2**63 - 1. Both values are unrepresentable in float64; the nearest float rounds up by 1 ULP, which then fails the dtype-range check.

Affected sites:

• xrspatial/geotiff/_geotags.py:619 (extract_geo_info — main GDAL_NODATA parse)
• xrspatial/geotiff/_reader.py:1166 (_resolve_masked_fill — LERC fill)
• xrspatial/geotiff/_reader.py:1299 (_sparse_fill_value — SPARSE_OK fill)

This is the same class of bug _vrt._parse_band_nodata (PR #1833) fixed for the VRT XML parse path: try int() first, fall back to float() only for NaN/Inf/scientific/fractional. The VRT-side fix did not extend to the TIFF source-of-truth parser, so:

• Direct open_geotiff(uint64.tif) with nodata=uint64 max → sentinel lost.
• write_vrt([uint64.tif]) stringifies the float-parsed sentinel into XML as 1.8446744073709552e+19, which the VRT reader subsequently rejects as out-of-range.

Categories: Cat 2 (NaN propagation: sentinel pixel survives unmasked) + Cat 5 (backend inconsistency: VRT XML path handles 64-bit sentinels via _parse_band_nodata but the TIFF path does not).

Fix sketch

Lift the int-first parse from _vrt._parse_band_nodata into a shared helper and reuse it across the three TIFF-side sites. The helper takes the raw string and the target dtype; it tries int(text), falls back to float(text) for NaN/Inf/fractional/scientific, and returns the parsed value at full Python int precision when possible.

Test plan

• uint64 nodata=264-1: NaN-mask fires, attrs['nodata'] surfaces 264-1 (int)
• int64 nodata=263-1: NaN-mask fires, attrs['nodata'] surfaces 263-1
• int64 nodata=-2**63 (representable): no regression
• uint16 nodata=65535 (fits float64): no regression
• uint64 source → write_vrt → read_vrt: sentinel round-trips