geotiff tests: consolidate parity tail cluster (Sub-PR B)

[brendancol]

Closes #2431 (Sub-PR B of two). Cluster 7 of the long-tail GeoTIFF test consolidation epic (#2424). Tests-only.

Summary

Fold eight top-level parity files into three focused siblings next to parity/test_backend_matrix.py, grouped by concern:

• parity/test_finalization.py -- read-finalization parity:
  • test_dispatch_validation_parity_2162.py
  • test_eager_finalization_parity_2162.py
  • test_lazy_finalization_parity_2162.py
• parity/test_signature_contract.py -- signature / docstring / release-contract parity:
  • test_signature_parity_1631.py
  • test_read_entry_points_doc_param_parity_2274.py
  • test_release_contract_parity_2389.py
• parity/test_reference.py -- degenerate-shape backend parity and external-reference round trips:
  • test_degenerate_shapes_backends_2026_05_11.py
  • test_round_trip_parity_rasterio_zarr_1961.py

backend_matrix.py is already 2090 lines; appending all eight files (~2578 lines) would push it past 4600. The issue allows a sibling once backend_matrix grows past 2500 lines. Three siblings keep each file well under that ceiling and group related tests together. The two signature-flavoured files (#1631, #2274) land with the release-contract gate so the contract-style parity files share a home; no test is duplicated against Sub-PR A.

GPU gating moves to the shared requires_gpu marker from _helpers/markers.py. The release-contract repo-root path moves from parents[3] to parents[4] to track the new directory depth. No assertion changed.

Audit

xrspatial/geotiff/tests/CLUSTER_AUDIT_PARITY.md maps every old file::test to its new file::test_id. The audit file is deleted in a final pre-merge commit.

Verification

• pytest xrspatial/geotiff/tests/parity/test_finalization.py xrspatial/geotiff/tests/parity/test_signature_contract.py xrspatial/geotiff/tests/parity/test_reference.py -q: 137 passed (same total as the eight source files).
• pytest xrspatial/geotiff/tests/parity/ -q: 575 passed, 43 skipped.
• pytest xrspatial/geotiff/tests/ --co -q: 5891 tests collected, no errors.

Test plan

• Three new parity/ files pass on this checkout (GPU present).
• Full xrspatial/geotiff/tests/ suite collects.
• CI green across numpy / cupy / dask+numpy / dask+cupy.
• Audit file deleted on the pre-merge commit.

[brendancol]

PR Review: geotiff tests consolidate parity tail cluster (Sub-PR B)

Tests-only consolidation. I read all three new parity files in full and compared them against the eight deleted source files. No source changed, so the domain-specific checks (algorithm accuracy, backend dispatch, cuda kernels) do not apply. The review is about whether the consolidation keeps coverage intact and whether the two non-trivial edits (the repo-root path depth change and the cross-module import) are correct.

Blockers

None.

Suggestions

None.

Nits

• parity/test_finalization.py defined _gpu_dask_available() below its first call site. It resolved fine at call time, but reading in definition order is cleaner. Fixed in 441e0f1 by hoisting the helper above the lazy-section fixtures.

What looks good

• Test count preserved exactly: 137 collected before and after (81 + 25 + 31 across the three siblings).
• Splitting into three siblings rather than one large append keeps every file well under the issue's 2500-line ceiling. backend_matrix.py would have hit roughly 4600 lines if everything landed there.
• The load-bearing edit checks out: test_signature_contract.py moves _REPO_ROOT from parents[3] to parents[4] because the file sits one directory deeper now. I verified the path resolves to the repo root and the release-contract section passes (25 passed).
• The cross-module import from ..read.test_crs import _write_rotated_tiff resolves and the lazy-finalization georef tests pass.
• GPU gating moved to the shared requires_gpu marker. The _BACKENDS parametrize marks and the inline conditional skips are both preserved via the runtime _gpu_dask_available probe.
• No duplicate top-level defs or classes in any file; the whole parity dir collects together cleanly (618 tests).
• isort and flake8 clean.

Checklist

• Test count preserved (137 -> 137)
• Repo-root path depth corrected for the new location
• Cross-module import resolves
• GPU gating preserved (requires_gpu + runtime probe)
• No dropped or duplicated tests; audit maps every old test
• Imports resolve; lint clean
• [n/a] Algorithm / cuda correctness (no source changed)
• [n/a] README feature matrix (tests-only)