geotiff tests: consolidate writer cluster (cog, bigtiff, overview)

[brendancol]

Closes #2400. Part of epic #2390 (PR 7 of 11).

Summary

Consolidates 26 writer-side test files into a four-file write/
subpackage:

• write/test_basic.py — generic writer paths (compression, tiling,
  kwarg order, return path, layout monkeypatch, VRT writer surface).
• write/test_cog.py — COG writer compliance and invalid-input
  errors.
• write/test_bigtiff.py — BigTIFF + COG compliance.
• write/test_overview.py — overview-level and nodata-aware overview
  tests.

Tests-only restructure. No source code changed.

Test ID parity

pytest --collect-only -q reports the same 497 test ids before and
after the move. Full geotiff suite (5706 tests) passes locally on
Python 3.14.

Module-level dedup

The fold caught a few shadowing collisions when several source files
defined identically-named module-level helpers:

• Identical _build_source_tif, _gpu_available, and
  _HAS_GPU = _gpu_available() lines: deduplicated, keeping the first
  occurrence.
• Two _float_da helpers with different shape defaults: the (8, 8)
  variant was renamed to _float_da_small so the later (64, 64) one
  no longer shadows it.
• The private from xrspatial.geotiff._vrt import write_vrt was
  aliased to _priv_write_vrt so it does not shadow the public
  write_vrt used by the signature-parity tests in the same module.

Doc-citation refresh

The release-gate checklist (docs/source/reference/release_gate_geotiff.rst)
and the user-facing GeoTIFF doc (docs/source/reference/geotiff.rst)
referenced several of the deleted filenames. Those rows were re-pointed
at the new write/*.py locations so the rst-parity gate
(test_release_gate_2321.py) keeps passing.

Audit deliverable

xrspatial/geotiff/tests/CLUSTER_AUDIT_PR7.md ships with this PR and
is deleted in a follow-up commit on the same branch before approval
per the epic contract.

Out of scope

• HTTP-side COG tests (integration cluster, PR 9).
• Reader-side files (PR 8).
• VRT non-writer files (PRs 1, 2, 4, 6).

Test plan

• pytest xrspatial/geotiff/tests/write/ -v
• pytest xrspatial/geotiff/tests/ -x -q
• CI (numpy / cupy / dask+numpy / dask+cupy) green

[brendancol]

PR Review: geotiff tests: consolidate writer cluster (cog, bigtiff, overview)

Blockers

• pytest.importorskip("rasterio") called at module scope mid-file in test_basic.py:314 and test_cog.py:410, 1396. When the source files were separate modules, each importorskip only skipped its own file. After folding, these calls still run at module-import time, so a single missing rasterio skips every prior test in the same module, including tests that do not need rasterio. Affected sections:

  • test_basic.py:314 skips folded sections from test_writer.py, test_writer_matrix.py, test_writer_kwarg_order_1922.py, and test_writer_return_path_1938.py if rasterio is absent.
  • test_cog.py:410 and :1396 skip earlier sections such as TestCOGWriter (which does not need rasterio) when rasterio is missing.

  Fix: replace each module-top pytest.importorskip("rasterio") with per-test calls inside the test functions that actually need rasterio, or apply a scoped pytestmark only to the affected sections.

Suggestions

• _helpers/markers.py:gpu_available is not reused. The epic states that markers come from _helpers/markers.py. test_overview.py:26-35 redefines _gpu_available() inline and test_cog.py:15 imports from ..conftest rather than from .._helpers.markers. Pointing all of them at the shared helper would prevent further drift.

• Redundant duplicate import lines in test_basic.py:31-48 and test_cog.py:28-34. The migration deduplicated identical full-line imports but did not collapse overlapping from xrspatial.geotiff import a, b, c / from xrspatial.geotiff import b, c, d pairs. Six lines all import to_geotiff and four all import write_vrt. The code works, but the noise will hide a future import drift.

• Comment-only references to deleted filenames still in source. xrspatial/geotiff/_writer.py:140 and xrspatial/geotiff/_attrs.py:339 still cite the deleted test_cog_requires_tiled_2312.py and test_bigtiff_cog_compliance_2286.py. Same for xrspatial/geotiff/tests/test_overview_block_order_2308.py:113 and xrspatial/geotiff/tests/test_release_gate_cog.py:12,16. No gate fails on these, but they read as broken pointers.

Nits

• # Folded from: <old_filename> banners still embed the issue numbers (e.g. # Folded from: test_writer_kwarg_order_1922.py). The epic says "drop issue numbers from filenames". The audit MD already captures the file map; the in-source banners could shrink to thematic names.

• test_basic.py:48 aliases write_vrt to _priv_write_vrt but the alias is single-use in one folded section. A brief inline comment next to the import would help a reader who arrives without the PR body.

• Module docstrings on the four new files list every folded source filename verbatim. That information lives in CLUSTER_AUDIT_PR7.md, which is itself temporary. Once the audit MD is removed before merge, the docstrings will be the only record. Fine either way; just calling out the duplication.

What looks good

• 497 test IDs collected on both sides of the move (verified with --collect-only -q).
• Both verification commands from the epic pass locally.
• The _priv_write_vrt aliasing is the right call: without it the private write_vrt (signature vrt_path, ...) shadowed the public one (signature path, ...) and broke the signature-parity test in the same module.
• Release-gate rst and geotiff.rst doc citations were updated together with the file move; the rst-parity gate (test_release_gate_2321.py) passes.

Checklist

• Algorithm matches reference/paper: N/A (tests-only)
• Backend parity: no source changes
• NaN handling: no source changes
• Edge cases preserved via verbatim fold
• Docstrings present on all four new files

[brendancol]

PR Review (follow-up after commit 23b45eb)

Disposition of original findings

Blocker

• Mid-module pytest.importorskip("rasterio") skip leakage. Fixed. write/test_basic.py:325 and the 8 affected tests in the write/test_cog.py compliance section now do rasterio = pytest.importorskip("rasterio") as the first statement in each test body. The compliance section's module-level binding was removed; the parity section's redundant module-top importorskip (whose tests already had per-test guards) was also removed. write/test_bigtiff.py keeps its module-top importorskip because that file is entirely rasterio-dependent (single source folded).

Suggestions

• Reuse _helpers/markers.py:gpu_available. Fixed. write/test_overview.py, write/test_basic.py, and write/test_cog.py now import gpu_available from .._helpers.markers instead of redefining it inline or routing through conftest.
• Collapse redundant import lines. Fixed. write/test_basic.py:31-48 and write/test_cog.py:28-34 are now one parenthesised from xrspatial.geotiff import (...) block each, plus the private write_vrt alias which carries an explanatory comment.
• Comment-only references to deleted filenames. Fixed. xrspatial/geotiff/_writer.py:140, xrspatial/geotiff/_attrs.py:339, xrspatial/geotiff/tests/test_overview_block_order_2308.py:113, and xrspatial/geotiff/tests/test_release_gate_cog.py:12,16 now point at write/test_cog.py / write/test_bigtiff.py.

Nits

• Section banner issue numbers. Fixed. # Folded from: test_*_NNNN.py banners renamed to # Section: <thematic name> (e.g. # Section: COG external-interop compliance suite).
• Comment next to the _priv_write_vrt alias. Fixed inline with the import-collapse: the alias now carries a short comment explaining the shadowing concern.
• Module docstrings re-list source filenames. Fixed. All four module docstrings now describe the topical coverage rather than re-listing the deleted source files.

Verification

pytest xrspatial/geotiff/tests/write/ -v : 497 passed.
pytest xrspatial/geotiff/tests/ -x -q : 5706 passed, 68 skipped, 6 xfailed, 1 xpassed.

No remaining findings from the first-pass review. CLUSTER_AUDIT_PR7.md is still in place per the epic; it gets deleted in a separate commit on this branch before merge.