Epic: Consolidate the GeoTIFF test suite

Background

xrspatial/geotiff/tests/ has grown to 354 test files / ~97k lines, plus one related file under xrspatial/tests/. The growth pattern is one file per GitHub issue: test_allow_rotated_geotiff_2115.py, test_backend_full_parity_2211.py, test_vrt_missing_sources_policy_1799.py, and so on.

The shape of the problem:

• Four backend-parity files, four attrs-contract files, three VRT-validation files, three rotated-CRS files. Each set covers nearly the same surface area with slightly different cases.
• Past "consolidated matrix" PRs landed without removing their predecessors. _1799 is a strict subset of _2367 but both still ship.
• File names read like a list of closed tickets. It is hard to look at the directory and see what is tested, what is duplicated, or what is missing.
• 28KB+ clusters where one parametrized test would do the same job.

Goal

Restructure xrspatial/geotiff/tests/ from ~354 files down to ~50, organised by concern. Drop issue numbers from filenames (git log and PR descriptions remain the audit trail). Tests-only restructure; no changes to xrspatial/geotiff/ source modules.

Target layout

xrspatial/geotiff/tests/
├── conftest.py                 (slim: fixtures + marker registration only)
├── _helpers/
│   ├── tiff_builders.py        (make_minimal_tiff, relocated from conftest)
│   ├── tiff_surgery.py         (relocated from current top-level)
│   └── markers.py              (requires_gpu, requires_loopback, requires_integration)
├── read/
│   ├── test_basic.py
│   ├── test_dtypes.py
│   ├── test_compression.py
│   ├── test_tiling.py
│   ├── test_endianness.py
│   ├── test_nodata.py
│   ├── test_crs.py             (rotated, dropped, missing, EPSG variants)
│   ├── test_coords.py
│   └── test_streaming.py
├── write/
│   ├── test_basic.py
│   ├── test_cog.py
│   ├── test_bigtiff.py
│   └── test_overview.py
├── vrt/
│   ├── test_validation.py
│   ├── test_missing_sources.py
│   ├── test_metadata.py
│   ├── test_window.py
│   └── test_dtype_conversion.py
├── attrs/
│   ├── test_contract.py
│   └── test_roundtrip.py
├── parity/
│   ├── test_backend_matrix.py
│   └── test_pixel_equality.py
├── release_gates/
│   └── test_stable_features.py
├── integration/
│   ├── test_http_sources.py
│   ├── test_dask_pipeline.py
│   └── test_gpu_pipeline.py
├── unit/
│   ├── test_header.py
│   ├── test_dtypes.py
│   ├── test_geotags.py
│   ├── test_safe_xml.py
│   └── test_compression.py
└── golden_corpus/              (unchanged; already well structured)

xrspatial/tests/test_geotiff_streaming_bigtiff_threshold_1785.py folds into geotiff/tests/read/test_streaming.py.

Conventions

• No issue numbers in filenames or test names. Git history is the trail.
• One file per concern. Each file owns a slice of the surface area and is the only place that slice is tested.
• Parametrize over backends × fixtures × variants with descriptive id= strings (id="rotated_no_crs[float32-deflate]", not id="bug_2126").
• make_minimal_tiff lives in _helpers/tiff_builders.py. conftest.py only holds fixtures and registers markers.
• Markers in one place: _helpers/markers.py (requires_gpu, requires_loopback, requires_integration). The pytest_collection_modifyitems socketserver hack in the current conftest.py is replaced by an explicit @requires_loopback marker on the affected tests.

Rollout

Each PR is self-contained. It creates the directories it needs, lands the consolidated tests, runs the audit (below), and deletes the superseded files in the same commit. The order goes from the smallest, lowest-risk cluster to the largest.

PR 1 — Foundation + VRT missing-sources

• Adds _helpers/ (tiff_builders, tiff_surgery, markers) and vrt/ directory.
• Collapses test_vrt_missing_sources_policy_1799.py + test_vrt_missing_sources_policy_2367.py into vrt/test_missing_sources.py.
• Smallest cluster; exercises the new layout end to end.

PR 2 — VRT validation

• test_vrt_validation_2321.py + test_vrt_capability_validator_2371.py + test_vrt_unsupported_2370.py → vrt/test_validation.py.

PR 3 — Rotated / dropped CRS

• test_allow_rotated_geotiff_2115.py + test_allow_rotated_crs_drop_2126.py + test_allow_rotated_no_crs_2122.py → read/test_crs.py, parametrized over (has_crs, has_rotation, drop_crs).

PR 4 — Backend parity

• test_backend_parity_matrix.py + test_backend_full_parity_2211.py + test_backend_pixel_parity_matrix_1813.py + test_attrs_kwarg_parity_1561.py → parity/test_backend_matrix.py + parity/test_pixel_equality.py.
• Largest cluster by lines; biggest reviewer load.

PR 5 — Attrs contract

• The four test_attrs_*_1984.py files (canonical, aliases, passthrough, version) → attrs/test_contract.py.

PR 6 — VRT metadata / window / dtype

• The ~120 remaining VRT-prefixed files split across vrt/test_metadata.py, vrt/test_window.py, vrt/test_dtype_conversion.py.
• May split into two sub-PRs if the review surface is too large.

PR 7 — Writer / COG / BigTIFF

• ~20 files → write/test_basic.py, write/test_cog.py, write/test_bigtiff.py, write/test_overview.py.

PR 8 — Reader paths

• Remaining read-side files → read/*.py (basic, dtypes, compression, tiling, endianness, nodata, coords, streaming).
• Folds in xrspatial/tests/test_geotiff_streaming_bigtiff_threshold_1785.py.

PR 9 — Accessor / integration

• ~20 files → integration/test_http_sources.py, integration/test_dask_pipeline.py, integration/test_gpu_pipeline.py.

PR 10 — Release-gate registry

• Every @pytest.mark.release_gate test pulled into release_gates/test_stable_features.py.

PR 11 — Unit-level cleanup

• Header / dtypes / geotags / safe_xml / compression unit tests → unit/*.py.
• Slim conftest.py to fixtures only. Drop the pytest_collection_modifyitems socketserver hack. Final lint pass.

PR title pattern: geotiff tests: consolidate <cluster> cluster. PR body links the superseded issue numbers so traceability stays in the GitHub record.

Audit method (per PR)

Each cluster PR adds a temporary CLUSTER_AUDIT.md that maps old test functions to their new home. It is deleted before merge.

A row may legitimately collapse: multiple old tests can map to one parametrized case. Dropping a row requires a justification in the Notes column. The audit table is the gate, rather than a coverage diff or trust in the parametrize design.

Reusing existing code

• make_minimal_tiff() (currently in conftest.py) already covers compression, tiling, geotransform, byte order, and BigTIFF. Relocate, do not rewrite.
• _tiff_surgery.py helpers move under _helpers/ unchanged.
• Golden corpus oracle (golden_corpus/_oracle.py) is already the right pattern; leave alone.
• Existing fixtures (simple_float32_tiff, simple_uint16_tiff, geo_tiff_data, tiled_tiff_data) stay in conftest.py.
• Markers keep their existing names (requires_gpu, requires_loopback, requires_integration).

Verification (per PR)

1. pytest xrspatial/geotiff/tests/ -x -q passes on a clean checkout.
2. pytest xrspatial/geotiff/tests/<new_cluster_dir>/ -v lists the expected parametrized IDs.
3. CLUSTER_AUDIT.md is filled in: every old file::test row is mapped or justified as dropped.
4. git diff --stat shows old cluster files deleted, not left behind.
5. pytest --collect-only xrspatial/geotiff/tests/ | wc -l drops, trending toward ~50 files over the full epic.
6. CI green across numpy / cupy / dask+numpy / dask+cupy backends, including release-gate jobs.

Definition of done

After PR 11:

• find xrspatial/geotiff/tests -name 'test_*.py' | wc -l is around 50.
• No filename matches test_*_[0-9]{4,}.py.
• pytest xrspatial/ -q is green across all backends.
• Coverage on xrspatial/geotiff/ is at or above the baseline captured before PR 1 lands.

Out of scope

• Changes to xrspatial/geotiff/ source modules.
• The golden_corpus/ subdirectory.
• Test suites outside xrspatial/geotiff/tests/, except the one BigTIFF threshold file noted above.