geotiff: parse GDAL_NODATA as int first to keep 64-bit sentinels exact (#1847)

[brendancol]

Summary

Closes #1847.

extract_geo_info, _resolve_masked_fill, and _sparse_fill_value parsed the GDAL_NODATA tag through float() unconditionally. uint64 max (2**64 - 1) and int64 max (2**63 - 1) are not exactly representable in float64; the nearest float sits one ULP above the dtype's max so the downstream info.min <= int(nodata) <= info.max gate rejected the cast and the sentinel pixel survived as a literal valid integer rather than being masked to NaN. write_vrt then stringified the float-parsed sentinel into the VRT XML as 1.8446744073709552e+19, which the VRT reader subsequently rejected as out-of-range -- so the bug propagated through the whole 64-bit pipeline rather than just the eager TIFF read.

The VRT XML parse path was already int-aware via _vrt._parse_band_nodata (PR #1833). This change lifts the int-first parse into a shared helper _parse_nodata_str in _geotags.py and routes the three TIFF-side sites through it. The helper tries int(text) first to preserve full precision, falls back to float(text) for NaN / Inf / scientific notation / fractional values. Downstream gates already handle integer values transparently because np.isfinite(int) works and int(int) is a no-op, so no call sites change.

Categories: Cat 2 (NaN propagation: sentinel pixel survived as a literal valid number on all four backends) + Cat 5 (backend inconsistency: VRT XML parse handled 64-bit sentinels but the TIFF parse fed by write_vrt did not).

Repro before fix

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

with tempfile.TemporaryDirectory() as d:
    arr = np.full((16, 16), 100, dtype=np.uint64)
    arr[0, 0] = 2**64 - 1
    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])
    # Before: uint64 18446744073709551615  (sentinel survives unmasked)
    # After:  float64 nan

Test plan

• Unit-level _parse_nodata_str matrix (int branch + float fallback + edge cases)
• Eager open_geotiff masks uint64 max / int64 max to NaN
• int64 min / uint16 max / int32 -9999 / float -9999.0 regression guards still pass
• read_geotiff_dask(chunks=) masks 64-bit sentinels via the windowed reader
• write_vrt -> read_vrt round-trip preserves the integer XML literal and masks the sentinel
• GPU path masks 64-bit sentinels (fix is upstream of per-backend code)
• All 2434 non-stale geotiff tests pass; 1 pre-existing failure unrelated (test_size_param_validation_gpu_vrt_1776.test_tile_size_positive_works asserts pre-to_geotiff accepts tile_size values that are not multiples of 16 #1767 tile_size=4 behaviour)

[Copilot]

Pull request overview

This PR fixes precision loss when parsing the GeoTIFF GDAL_NODATA tag for 64-bit integer sentinels by parsing as int first (preserving exact uint64/int64 max), with a float fallback for NaN/Inf/scientific/fractional values. This aligns the TIFF-side nodata parsing with the existing VRT XML path and prevents downstream masking + VRT round-trip failures for 64-bit nodata.

Changes:

• Add shared _parse_nodata_str helper in _geotags.py and use it in extract_geo_info.
• Route _reader._resolve_masked_fill and _reader._sparse_fill_value through the shared int-first nodata parser.
• Add a comprehensive regression test suite covering eager, dask, VRT round-trip, and GPU paths for 64-bit nodata sentinels.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

File	Description
xrspatial/geotiff/_geotags.py	Introduces _parse_nodata_str and uses it for GDAL_NODATA parsing in extract_geo_info.
xrspatial/geotiff/_reader.py	Updates LERC masked-fill and sparse-fill nodata parsing to use the shared helper.
xrspatial/geotiff/tests/test_nodata_int64_precision_1847.py	Adds regression coverage for uint64/int64 sentinel masking + VRT round-trips + dask/GPU parity.
.claude/sweep-accuracy-state.csv	Updates audit tracking entry for the geotiff module and this fix.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[brendancol]

Addressed all three: unused tempfile import removed, GPU gating switched to the _gpu_only / _HAS_GPU helper used by test_gpu_nodata_1542.py, and _parse_nodata_str signature widened to str | None. I also widened GeoInfo.nodata from float | None to int | float | None since that's the one downstream type that now actually receives int values from the helper. The _resolve_masked_fill / _sparse_fill_value callers don't carry return annotations and their nodata_str parameter is already str | None, so no further churn needed there.