[jmschrei]

This is the second of three test-only PRs from a file-by-file review of the suite. It fixes or removes tests that could not fail, and makes three tests that never ran start running. The next PR removes redundant tests and merges duplicates; the one after expands weak assertions.

How each change was checked. Every check added or tightened here was run against a copy of the library broken in the way the check guards, applied in a scratch copy or by monkeypatching, never in the repository. Each failed there and passes on the real code. Every removed test was read against the test said to cover it.

Tests that passed whatever the code did

• test_match: test_extract_and_filter_chrom_signal_threshold tested thresholds 30 and 10.
  • At 30 every window passes, the same as with no bigWig. At 10 none do, so its second half asserted inside a loop over {}.
  • It now runs 10, 15, 21 and 30, and checks how many windows are kept, the empty case included.
  • The GC-window tests caught any error from extract_loci with a bare except: continue. The call is no longer wrapped.
• test_saturation_mutagenesis:
  • test_saturation_mutagenesis_ordering compared at 2 decimals attributions that are all below 0.0025, so it passed against zeros. It now compares at 4.
  • test_saturation_mutagenesis_packed never reached the fast path, because a job this small skips it. It now forces the fast path and fails if it falls back.
• test_deep_lift_shap:
  • test_deep_lift_shap_random_state compared 16 examples with 10, so it passed on the shape mismatch alone.
  • A missing comma made the numpy-multipliers case of test_hypothetical_attributions_raises repeat another case. The "Shape of all tensors must match." branch was not tested; it now is.
  • The cleanup-on-error tests raised on target=5 before any backward hook ran, so they passed with a cleanup that only removes handles. The hook test now raises inside the softmax rule, and both check every attribute _clear_hooks removes.
  • Three tests asserted the dtype of the input fixture instead of the attributions'.
• test_pisa:
  • test_pisa_n_shuffles and three "tiny threshold warns" lines called deep_lift_shap instead of pisa.
  • In test_pisa_raises, the references came from a two-sequence fixture while X had sixteen sequences, so the channel and length lines failed on the batch axis. They are now built from the test's own X and matched to their shapes.
• test_ersatz:
  • Three lines in each of the insert and substitute raise_ends tests passed an int as alphabet.
  • The raises_alphabet motifs had no batch axis, so they never reached the channel or length check.
  • test_local_dinucleotide_shuffle_bin_layout_valid checked a helper defined in the test file. It now checks the library's _local_bins over 200 seeds.
• design:
  • test_beam_substitution_max_iter_default and _input_mask passed when no search step ran.
  • test_greedy_substitution_reverse_complement never set reverse_complement=True. It now compares that search with one over the motifs and their reverse complements listed explicitly.
• test_plot: test_plot_logo_start_end_slicing passed with start and end ignored. It now compares the result with the logo of the sliced attributions.

Checks of pandas, figwig, torch, the toy models, or a call against itself

• Removed tests:
  • test_load_signals_nan (figwig).
  • test_predict_raises_shape (the toy model's permute).
  • test_marginalize_additional_func_kwargs_none (the default passed explicitly).
  • test_marginalize_annotations_threads_kwargs: equal outputs whether or not the keywords are forwarded.
  • test_saturation_mutagenesis_no_warning_on_hard_one_hot: the same call as two other tests.
  • test_plot_logo_color_dict: only ax is not None, which _still_per_character covers.
  • test_transformer_block_attention_mask_forward (toy model only).
• Removed lines and cases:
  • the df.__eq__ lines in the _interleave_loci tests (pandas);
  • the n_jobs=1 case of test_extract_loci_n_jobs;
  • the batch sizes that are clamped to the same single batch in test_dependency_map_batch_size and test_saliency_batch_size;
  • the torch Conv1d line of test_screen_raises.
• Unused device fixture removed from tests that raise before anything moves. Their CUDA cases repeated the CPU ones.
• deep_lift_shap rule tests: test_deep_lift_shap_rules_input_type (45 calls on identical inputs) and test_deep_lift_shap_rules_return_references (a NamedTuple field compared with itself) are removed. Their one real check each moves into test_deep_lift_shap_input_type and test_deep_lift_shap_return_references_named_tuple.
• Attention mask: a forward pass now checks that the toy model's mask changes its output. This replaces eight deep_lift_shap calls.

Tests that never ran

• Without BioPython. test_extract_loci_fasta_bgzip_without_biopython checks the error pyfaidx raises for a block-gzip FASTA without BioPython, and was skipped wherever BioPython was installed. pyfaidx imports BioPython only when it opens such a file, so the test now hides it through sys.modules and runs either way. Its counterpart, test_extract_loci_fasta_bgzip, still runs only where BioPython is installed, which CI does not install.
• Console script. test_console_script_print_path is marked cmd, which addopts and CI deselect. Three new tests call main() in process instead: --print-path, --dest, and a repeated --dest with and without --force.
• Skipped cases. Four cases of test_packed_sequence_from_one_hot_raises_values skipped. The parameters are now explicit (dtype, value) pairs.
• Skipped test. test_deep_lift_shap_subclassed_op is now a strict xfail instead of a skip.

Checks

• Full suite on the CPU and one CUDA device: 3,504 passed, 3 skipped, 2 xfailed. The skips are pisa's two existing CUDA TODOs and the block-gzip FASTA test, which needs BioPython.
• sphinx -W passes.
• prepush's credits check flags @RULE_PARAMS, a removed decorator, as a removed handle. Nothing was credited.

The six surfaces

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

🤖 Generated with Claude Code