[jmschrei]

This is the third test-only PR from the file-by-file review of the suite. It removes tests that another test fully covers and merges tests that differ only in an input or parameter into parametrized tests. Tests that overlap partially stay. The last PR in the series adds value checks to weak tests.

How each change was checked

• Removals. For each removed test, the library was broken in a way that test catches. The break was made in a scratch copy, or by monkeypatching from a scratch plugin, never in the repository. The test named as covering it then had to fail too. Where it did not, the test stayed:
  • test_captum_deep_lift_shap_n_shuffles stays, because it is the only test where n_shuffles is smaller than the number of references given.
  • test_predict_flattendense_fp16 stays, because its pinned values fall outside the tolerance of test_predict_flattendense_16bit. It is merged with its bf16 twin instead.
  • test_deep_lift_shap_invalid_target stays, because TEST fix or remove tests that could not fail, and run three that never ran #128 changed the test that used to overlap it.
• Merges. Each merge kept every assertion of the tests it replaced, for the same inputs. The assertions were listed before and after the edit. Pinned arrays were copied as source text and checked equal by parsing.
  • Where the originals checked different things, the merged test checks both for every case wherever that holds.
  • Several merges therefore gain a check, for example the dtype in both hypothetical-attribution cases and the value check for numpy annotations.
  • In test_ablate_deep_lift_shap_alpha, the beta case now checks that the attributions scale by beta, not only that they differ.
• Skipped findings. Findings the review rated low confidence are not in this PR.

Removed or merged, by file

• test_seqlet: fifteen older recursive_seqlets tests are removed, because the brute-force reference, error, tensor and control-threshold tests that UPDATE recursive_seqlets to decode every control threshold in one pass per row #125 added cover them:

  • numpy input;
  • the return_fdr and empty-control errors;
  • the default thresholds;
  • an out-of-range control;
  • an unreachable target;
  • control dtypes;
  • the p-value oracle, own-span p-values and sorting;
  • coordinate invariants;
  • flanks.

  The empty-X and invalid-threshold cases join test_recursive_seqlets_errors with their messages. Four tests fold into the tests that made the same calls.

• test_deep_lift_shap:

  • The 20 one-model Sequential tests become one table, and so do the 11 max-pool tests (9 1D, 2 2D) and the captum comparisons.
  • Tests that ran the same call once for convergence and once for pinned values (the rule models, the transformer blocks) now run it once.
  • The attention, shared-module and cleanup tests are parametrized over the model.
  • Removed: LayerNorm/RMSNorm (rows of the rules table, with the same pins); positional-is-linear; two _disable_hooks tests; the ConvDense dense wrapper (its attributions are bitwise equal to FlattenDense's); custom_op_overrides_builtin (its loop never passed the custom op).

• test_pisa:

  • The 20 one-model tests become test_pisa_sequential.
  • The rules convergence and regression tests make one call.
  • test_pisa_references_ignores_random_state is removed.
  • The pisa test named test_deep_lift_shap_shuffle_ordering is renamed.

• test_predict:

  • The nine fp32 architecture tests and the fp16/bf16 pairs become parametrized tests.
  • The ConvDense batch-size tests compare the outputs by value.
  • _flattendense_fp64 is removed; test_predict_stored_dtype[float64] covers it.

• test_saturation_mutagenesis: start/end, identity, irregular-column, input-dtype (now exact), span-error and low-precision tests are parametrized.

• test_io:

  • Parametrized: _interleave_loci on paths and DataFrames, the summit errors (now matched to their messages), and the out_window, alphabet, integer-chromosome, width and count-threshold cases. The thresholds are now checked as inclusive (>=, <=).
  • Removed: a duplicated summits test, drops_sample_columns, seq_first_n, load_signals_values and one_hot_to_fasta_packed.

• test_utils: the hand-written one_hot_encode cases, alphabet containers, characters cases, PackedSequence dtype and indexing tests, and packed random_one_hot are parametrized. Five one_hot_encode tests that reference comparisons cover are removed.

• test_match: three GC-window tests are cases of the multiples-of-5 test, so they are removed. The determinism test repeated a call whose output test_extract_matching_loci pins, so it is removed too. The _calculate_char_perc and _extract_and_filter_chrom cases are parametrized, the latter now comparing whole dicts.

• Smaller files: ablate, product, marginalize, space, variant_effect, saliency, dependency_map, annotate, kmers, plot and design/test_screen merge tests that differ only in the function, dtype, container or keyword.

  • Batch size 68 is clamped to 64. Two shape-only tests that ran at 68 are removed, and the ConvDense test now compares batch size 68 by value.
  • variant_effect's device test now also covers substitution_effect.

Size

	Before (#128)	After
Test functions	1,363	1,107
Collected cases	3,509	3,405
Lines in test files	28,371	25,436

Checks

• Full suite on the CPU and one CUDA device: 3,400 passed, 3 skipped, 2 xfailed. The skips are pisa's two existing CUDA TODOs and the block-gzip FASTA test, which needs BioPython.
• The CUDA cases of every merged file were also run on their own.
• sphinx -W passes.
• prepush flags @RULE_PARAMS and @TRANSFORMER_PARAMS, which are removed decorators, as removed credits. It also notes one space-indented line, copied unchanged from a pinned array in the original test.

The six surfaces

1. Tests: they are the change.
2. Docs, README: not touched.
3. Changelog: one bullet under Testing.
4. Bundled skill: not affected.
5. Compatibility: no library code changes.

🤖 Generated with Claude Code