geotiff: extend #1752 size-arg validation to write_geotiff_gpu, read_geotiff_gpu, read_vrt #1776

What's broken

Issue #1752 added tile_size and chunks validation to to_geotiff and read_geotiff_dask. The matching kwargs on three sibling entry points were skipped, so they still reach the deeper writer/dask paths with invalid values and surface confusing errors that do not name the bad parameter.

The gaps:

1. write_geotiff_gpu(tile_size=) is missing the validation to_geotiff does. The four cases geotiff: validate tile_size and dask chunks at entry points #1752 covers for to_geotiff:

   • tile_size=0 raises ZeroDivisionError from gpu_compress_tiles
   • tile_size=-1 raises struct.error ('H' format requires 0 <= number <= 65535)
   • tile_size=False raises ZeroDivisionError (bool slipped through because False == 0)
   • tile_size=256.0 raises TypeError ('float' object cannot be interpreted as an integer)

2. read_geotiff_gpu(chunks=) is missing the validation read_geotiff_dask does:

   • chunks=0 raises ZeroDivisionError
   • chunks=-1 silently produces output (no validation at all)
   • chunks=(0, 256) raises ZeroDivisionError
   • chunks=(256, -1) silently produces output

3. read_vrt(chunks=) is missing the same validation as 2.

open_geotiff(chunks=) is fine on the CPU path because it routes through read_geotiff_dask's already-validated branch. The gap shows up for open_geotiff(gpu=True, chunks=X), direct read_geotiff_gpu(chunks=X), and read_vrt(chunks=X).

How to reproduce

import numpy as np
import xarray as xr
import cupy
import os
import tempfile
from xrspatial.geotiff import (
    write_geotiff_gpu, read_geotiff_gpu, read_vrt,
    to_geotiff, write_vrt,
)

arr = np.arange(100, dtype=np.float32).reshape(10, 10)
da = xr.DataArray(arr, dims=['y', 'x'])

with tempfile.TemporaryDirectory() as d:
    tif = os.path.join(d, 'a.tif')
    to_geotiff(da, tif)
    vrt = os.path.join(d, 'a.vrt')
    write_vrt(vrt, [tif])

    write_geotiff_gpu(xr.DataArray(cupy.asarray(arr), dims=['y', 'x']),
                      os.path.join(d, 'out.tif'),
                      tile_size=0)        # ZeroDivisionError
    read_geotiff_gpu(tif, chunks=0)        # ZeroDivisionError
    read_geotiff_gpu(tif, chunks=-1)       # silently passes
    read_vrt(vrt, chunks=-1)               # silently passes

Suggested fix

Mirror the up-front guard #1752 added to to_geotiff and read_geotiff_dask:

• write_geotiff_gpu: reject non-positive int or non-int tile_size with a ValueError naming the parameter.
• read_geotiff_gpu and read_vrt: reject non-positive int or non-int chunks (and tuple-of-non-positive, and wrong-length tuple) with the same message format read_geotiff_dask uses.

Source touches are small; the guards already exist on the sibling entry points and can be reused.

Tests

Tests that should be added in the same PR:

• tile_size=0, tile_size=-1, tile_size=256.0, tile_size=True/False on write_geotiff_gpu
• chunks=0, chunks=-1, chunks=(0, N), chunks=(N, -1), wrong-length tuple, non-int on read_geotiff_gpu and read_vrt
• Positive int and np.int64 scalar still work (regression pin against the integer-scalar branch).

Surfaced by the test-coverage sweep on the geotiff module (pass 13, 2026-05-13).