PR Review: Make corpus determinism check toolchain-agnostic

Blockers

None.

Suggestions

• _assert_semantic_equal only compares base-IFD pixels even when both files have an overview chain (test_corpus_determinism.py:143-154). For the COG fixture, an overview-pixel regression in the generator would slip past: the decimation factor list matches but the overview pixels never get read. Consider iterating the overview levels for lossless cells with rasterio.open(path, OVERVIEW_LEVEL=lvl) and comparing each level's pixel array. _oracle.py already uses that pattern in compare_to_oracle.
• Lossy JPEG cells skip pixel comparison entirely (line 147-148). A pixel-content regression in the generator's JPEG branch (wrong input, wrong band order before encode) would not be caught here. A coarse band-wise mean check within a loose tolerance would close the gap without re-introducing libjpeg coupling.

Nits

• _is_encoder_sensitive, _assert_semantic_equal, and _nodata_equal have no direct unit tests. The drift-detection paths were verified manually in the PR; a tiny test that builds a doctored copy with a flipped pixel and asserts _assert_semantic_equal raises would lock the negative path in.
• The tmp-dir name in regenerated_dir is still "regen_corpus_1930" (line 178). Harmless but misleading now that the test's scope was extended in #2299.

What looks good

• Dispatch is driven from manifest fields (cog, compression, tolerance.lossy), not hard-coded fixture ids, so new encoder-sensitive cells inherit the right comparison mode automatically.
• The semantic check covers dtype, shape, transform, CRS, nodata (NaN-aware via _nodata_equal), overview decimation chain, and lossless pixels.
• The workflow comment explains why the --ignore is gone instead of leaving a silent loosening.

Checklist

• Algorithm matches reference: no algorithm; test-infra change
• All implemented backends produce consistent results: rasterio-only
• NaN handling is correct: _nodata_equal + equal_nan=True
• Edge cases: missing-fixture skip preserved
• Overview-level pixels exercised: not yet (see suggestion)
• Dask chunk boundaries: n/a
• No premature materialization: n/a
• Benchmark: n/a
• README feature matrix: n/a
• Docstrings present and accurate

Follow-up review

All four findings from the prior review pass were addressed in f5aa2141:

• Suggestion 1 (overview-level pixels): _assert_semantic_equal now iterates rasterio.open(..., OVERVIEW_LEVEL=lvl) for every level the file declares and compares each level's pixel array (lossless cells). See _assert_overview_pixels_exact at test_corpus_determinism.py:194.
• Suggestion 2 (lossy mean tolerance): _assert_pixels_close_lossy at test_corpus_determinism.py:209 computes per-band means and rejects drift above _LOSSY_PIXEL_MEAN_TOL = 4.0 (0-255 scale).
• Nit 1 (drift unit tests): two new tests cover the negative path: test_semantic_equal_rejects_lossless_pixel_drift and test_semantic_equal_rejects_lossy_mean_drift (test_corpus_determinism.py:357 and :371).
• Nit 2 (tmp-dir rename): tmp_path_factory.mktemp("regen_corpus_determinism") at test_corpus_determinism.py:256.

The branch was rebased onto origin/main after the follow-up so the PR diff is now scoped to the two #2299 files; force-push was clean.

Local run: pytest xrspatial/geotiff/tests/golden_corpus/: 96 passed, 1 skipped.

No new blockers, suggestions, or nits from this pass.