[brendancol]

Closes #2446. PR A of cluster #2425; part of the long-tail epic #2424.

Summary

Folds 9 top-level codec files into a new unit/test_codec_roundtrip.py.
The sibling unit/test_compression.py keeps its PackBits-kernel-only
scope (its docstring already says round-trips live elsewhere), so the
new file is a sibling rather than an extension.

Files folded:

• test_jpeg.py
• test_jpeg2000.py
• test_lerc.py
• test_lerc_max_z_error.py
• test_lerc_valid_mask.py
• test_lz4.py
• test_lz4_compression_level_2026_05_11.py
• test_compression_docstring_1644.py
• test_compression_level.py

Net file delta: 9 deleted, 1 added => -8 files.

What is preserved

• 112 tests in scope, all green locally.
• LERC max_z_error numerical tolerances pinned verbatim
  (budget + 1e-7 slack, bit-exact lossless via assert_array_equal).
• LERC and JPEG 2000 optional-dep gating via pytest.importorskip /
  skip markers.
• The to_geotiff(compression='jpeg') rejection contract.
• The rasterio-only GDAL tiled JPEG read-back tests, still skip-gated.
• The write_geotiff_gpu docstring drift guard from geotiff: write_geotiff_gpu compression docstring lists 4 codecs but accepts the full 9 #1644.

What changed shape

• Parametrise dtype / tiled / level / mask matrices with descriptive
  ids (e.g. jpeg2000[uint16-tiled], lerc[float32-stripped],
  lz4-level[16], lerc-mask[uint16-sentinel]).
• No issue numbers in test or filenames.
• LERC max_z_error classes flattened into a single class.
• LERC valid-mask round-trip variants parametrised over
  dtype × nodata × invalid_positions.

Out of scope (PR B, parallel)

PR B owns the photometric cluster and consolidates these into a
different unit/test_photometric.py:

• test_miniswhite_nodata_1809.py
• test_miniswhite_writer_roundtrip_1836.py
• test_signed_miniswhite_rejected_2278.py
• test_to_geotiff_allow_internal_only_jpeg_parity.py

GPU codec files (test_jpeg_gpu_*, test_lerc_valid_mask_gpu.py,
test_nvcomp_*, test_nvjpeg*) are scoped to the GPU cluster issue
#2438.

Audit trail

xrspatial/geotiff/tests/CLUSTER_AUDIT_CODEC_ROUNDTRIP.md maps every
old file::test to its new file::test_id. It lives on this branch
during review and is deleted in a final pre-merge commit so it does
not land on main.

Test plan

• pytest xrspatial/geotiff/tests/unit/test_codec_roundtrip.py -v
  (112 passed)
• pytest xrspatial/geotiff/tests/unit/ -q (247 passed)
• pytest xrspatial/geotiff/tests/ -x -q (5778 passed, 68 skipped,
  6 xfailed)
• find xrspatial/geotiff/tests -name 'test_*.py' | wc -l drops by
  8 (259 -> 251)

[brendancol]

PR Review: geotiff tests: consolidate codec round-trip cluster

Tests-only change. No edits to xrspatial/geotiff/ source modules. 9 top-level files folded into a new unit/test_codec_roundtrip.py. 112 tests preserved (same baseline count).

Blockers

None.

Suggestions

• xrspatial/geotiff/tests/unit/test_codec_roundtrip.py:1308-1315: the comment for _GPU_FALLBACK_CODECS now says "the JPEG round-trip is covered with appropriate uint8 RGB data elsewhere" without naming where. The original file pointed at test_gpu_writer_compression_modes_2026_05_11.py. Since GPU codec consolidation is owned by #2438, the most accurate phrasing here is "covered by the GPU codec cluster (issue #2438)". One-line wording tweak.
• xrspatial/geotiff/tests/unit/test_codec_roundtrip.py:96-100 and :475-481: there are two helpers named _make_compressible and _make_compressible_large with slightly different shapes ((128,128) vs (512,512)) and noise levels. Keeping them separate is correct (the lz4 size-effect test uses a smaller surface than the zstd/deflate one for runtime reasons), but the naming is easy to misread. Renaming the module-level helper to _make_compressible_smooth_small or inlining it into TestLz4CompressionLevel would remove the ambiguity.
• xrspatial/geotiff/tests/unit/test_codec_roundtrip.py:617-626: the LERC mask predicate closure captures invalid_positions via the default-arg pattern (def invalid_pred(a, positions=invalid_positions)). Good defensive choice. Worth a one-line comment explaining the bound-default pattern is intentional; without it a future refactor could collapse it to a free closure and reintroduce the late-binding bug.

Nits

• xrspatial/geotiff/tests/unit/test_codec_roundtrip.py:53: from xrspatial.geotiff._compression import (...) is a long parenthesised import. The leading items COMPRESSION_JPEG, COMPRESSION_JPEG2000, COMPRESSION_LERC, COMPRESSION_LZ4 are constants and the rest are functions; isort already keeps them in one block but a blank-line break after the constants would make the import easier to read.
• xrspatial/geotiff/tests/unit/test_codec_roundtrip.py:200-204: TestPublicToGeotiffJpegRejected is a class with a single test method. Both the original and the consolidated version use a class; nothing wrong with it, just notable.
• xrspatial/geotiff/tests/unit/test_codec_roundtrip.py:1098, :1110: path = str(tmp_path / "lz4_bad.tif") is shared across the parametrised cases. pytest's tmp_path is unique per test invocation, so there is no real collision risk, but the project memory note recommends including unique tokens in tmp filenames. Either include the level in the filename (e.g. f"lz4_bad_{level}.tif") or leave a comment saying tmp_path already provides isolation.
• CLUSTER_AUDIT_CODEC_ROUNDTRIP.md: scheduled for deletion on the pre-merge commit. The file header already says so; flagging here only so the deletion does not get forgotten.

What looks good

• LERC max_z_error numerical tolerances preserved verbatim (budget + 1e-7 slack, bit-exact lossless via assert_array_equal).
• pytest.importorskip / module-level skip markers kept for the optional codec libraries (glymur, lerc, lz4, rasterio).
• Parametrise ids are descriptive and free of issue numbers (lerc[float32-stripped], lz4-level[16], lerc-mask[uint16-sentinel]).
• The to_geotiff(compression='jpeg') rejection contract is still pinned (TestPublicToGeotiffJpegRejected).
• The rasterio-only GDAL tiled JPEG class is still skip-gated at the class level, so the non-rasterio environment still exercises the codec-only tests above it.
• The lerc_writer_with_mask monkeypatch fixture is preserved with the same predicate-callable indirection.
• Audit file maps every old file::test to the new file::test_id.

Checklist

• Tests-only; no source changes
• All implemented backends produce consistent results
• NaN handling is correct (LERC NaN-nodata case still pinned)
• Edge cases are covered by tests (boundary levels, empty payloads, single pixel)
• Optional-dependency skip markers preserved
• No premature materialization or unnecessary copies
• No benchmark needed (test consolidation only)
• No README feature matrix change needed
• Docstrings present on every consolidated class

[brendancol]

Follow-up review after nit pass

Disposition of original findings:

Suggestions

• GPU codec referral in _GPU_FALLBACK_CODECS comment: fixed in 26657e5. Now names the GPU cluster issue #2438 instead of pointing at a file that the GPU cluster will rename.
• _make_compressible / _make_compressible_large ambiguity: fixed in 26657e5. Module-level helper renamed to _make_compressible_smooth_small; docstring explains the size split.
• Default-arg late-binding guard in invalid_pred: fixed in 26657e5. Comment added explaining the bind is intentional.

Nits

• _compression import grouping: fixed in 26657e5. Section comments separate constants / availability flags / functions.
• TestPublicToGeotiffJpegRejected single-method class: dismissed. The original module used a class wrapper here too; flipping to a free function is stylistic preference, not a fix.
• Unique tmp_path filenames for parametrise: fixed in 26657e5. lz4_bad.tif, lz4_dask_bad.tif, and the generic bad.tif now include codec/level tokens.
• CLUSTER_AUDIT_CODEC_ROUNDTRIP.md deletion reminder: tracked. Will be removed in the pre-merge commit on this branch.

Other notes

• Merged origin/main (commit fe16afe) without conflicts. The merge only added an unrelated nodata-rejection test file at the top level.
• 112 tests still pass post-merge and post-fixes.

No new findings. Ready for the pre-merge audit-file deletion commit.