Fix read_geotiff_dask losing nodata mask in non-first chunks (#1597)

[brendancol]

Closes #1597.

Summary

• For integer rasters with an in-range nodata sentinel, read_geotiff_dask declared the dask array as float64 but only promoted individual chunks to float64 when the chunk actually contained a sentinel pixel. Chunks without a sentinel stayed at the file's integer dtype.
• Dask concatenation preallocates the result buffer from the first chunk's actual dtype, so a sentinel landing in any non-first chunk caused the compute result to come back as integer with the would-be NaNs cast to 0 (and a RuntimeWarning about an invalid cast).
• Threading the resolved target_dtype unconditionally through _delayed_read_window keeps every chunk's dtype consistent with the dask graph. Out-of-range sentinels still keep the file's native dtype because effective_dtype short-circuits that case earlier in read_geotiff_dask.

Test plan

• New regression suite test_dask_int_nodata_chunks_1597.py covering the sentinel-in-corner case, per-chunk dtype uniformity, the out-of-range Reading uint TIFF with negative GDAL_NODATA sentinel raises OverflowError #1581 path, and float-raster parity.
• Existing tests still pass: test_nodata_out_of_range_1581, test_attrs_parity_1548, test_vrt_int_nodata_1564, test_metadata_round_trip_1484, test_dtype_read, plus 657 non-GPU geotiff tests.

[Copilot]

Pull request overview

This PR fixes a correctness issue in the GeoTIFF dask reader where nodata masking could be silently lost when an in-range integer nodata sentinel only appears in non-first chunks, causing dask concatenation to coerce later chunks and convert NaN to 0.

Changes:

• Thread the resolved target_dtype through each delayed chunk read to keep per-chunk dtypes consistent with the dask array’s declared dtype.
• Add a regression test suite covering the non-first-chunk sentinel case, per-chunk dtype uniformity, out-of-range sentinel behavior, and float-raster parity.
• Update sweep metadata to track issue #1597 inspection state.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
xrspatial/geotiff/__init__.py	Ensures dask-delayed chunk reads receive the resolved target_dtype to prevent per-chunk dtype divergence that breaks nodata masking.
xrspatial/geotiff/tests/test_dask_int_nodata_chunks_1597.py	Adds regression coverage for the dtype/nodata issue when the sentinel is only present in non-first chunks.
.claude/sweep-metadata-state.csv	Updates internal sweep metadata to reference issue #1597.

Comments suppressed due to low confidence (1)

xrspatial/geotiff/init.py:1665

• Passing target_dtype unconditionally means _delayed_read_window will now always run arr.astype(target_dtype) for every chunk (because target_dtype is never None). Since astype copies even when the dtype is already correct, this can add a full extra copy per chunk for the common case where no dtype promotion is needed (e.g. no nodata masking, float rasters, or integer rasters without an in-range sentinel). Consider updating _delayed_read_window to only cast when arr.dtype != target_dtype (or use astype(..., copy=False)), so #1597 is fixed without introducing a per-chunk copy/perf regression.

            block = da.from_delayed(
                _delayed_read_window(source,
                                     r0 + win_r0, c0 + win_c0,
                                     r1 + win_r0, c1 + win_c0,
                                     overview_level, nodata,
                                     band_arg,
                                     target_dtype=target_dtype,
                                     http_meta_key=http_meta_key,
                                     max_pixels=max_pixels),
                shape=block_shape,
                dtype=target_dtype,
            )

---

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