Add Hypothesis property-based tests, fix 3 bugs found

[aarmey]

Summary

Adds ddmc/tests/test_properties.py: property-based tests (via Hypothesis) for the core sequence-distance and preprocessing helpers in ddmc.pam250, ddmc.binomial, ddmc.motifs, and ddmc.logistic_regression. Instead of hand-picked examples, each test generates many random-but-valid inputs and checks an invariant that should hold for all of them (symmetry, round-tripping, matching a reference implementation, etc.).

Adds hypothesis to the dev dependency group.

Running these tests turned up three real bugs, all now fixed in this PR:

1. get_pam250_scores silently overflowed int8. The pairwise PAM250 similarity matrix was accumulated into an int8 array. A length-11 peptide's self-similarity score can exceed int8's 127-value ceiling for highly self-similar residues — e.g. 11 tryptophans score 11 * 17 = 187, which wrapped around to -69. This meant highly conserved/repetitive motifs (poly-W, poly-C) could be scored as less similar to themselves than to unrelated sequences, silently corrupting the PAM250 distance metric used by DDMC(distance_method="PAM250").

   >>> get_pam250_scores(["WWWWWWWWWWW", "CCCCCCCCCCC"])
   array([[ -69,  -88],
          [ -88, -124]])   # before the fix

   Fix: widen the accumulator from int8 to int32 (ddmc/pam250.py).

2. compute_control_pssm crashed on the exact input its real caller produces. It looked up each residue with AAlist.index(aa), which raises ValueError for a lowercase letter. But ddmc.binomial.BackgroundSeqs — the only place compute_control_pssm is actually called from, via DDMC.get_pssms(PsP_background=True) (ddmc/clustering.py:281-282) — always lowercases the phosphoacceptor of the sequences it returns. So DDMC.get_pssms(PsP_background=True) was broken end-to-end, raising ValueError: 'y' is not in list (or 's'/'t').
   Fix: upper-case each residue before the AAlist lookup (ddmc/motifs.py).

3. normalize_cluster_centers centered along the wrong axis. Its docstring says it zero-means each cluster (column) across samples (rows) — the usual per-feature standardization before classification. But the implementation wrapped StandardScaler(with_std=False).fit_transform in centers.T ... .T, which instead zero-meaned each sample across that sample's clusters, leaving the per-cluster mean across samples unchanged (and generally nonzero).
   Fix: call StandardScaler(with_std=False).fit_transform(centers) directly, without the transposes (ddmc/logistic_regression.py).

The other seven property tests passed from the start and pin down invariants worth protecting going forward — fast_position_weight_matrix numerically matching the Biopython reference PWM, GenerateBinarySeqID's one-hot encoding round-tripping to the original sequence, CountPsiteTypes accounting for every input sequence, and PAM250-score symmetry. All ten tests, including the three that exposed bugs, now pass as ordinary regression tests (no xfail markers needed).

Test plan

• uv run pytest ddmc/tests/test_properties.py -v — 10 passed
• uv run pytest ddmc/tests/ -q --deselect ddmc/tests/test_datasets.py — 54 passed (no regressions; test_datasets.py deselected as it needs network-fetched data not available in this environment)
• uv run ruff check ddmc/ — clean

🤖 Generated with Claude Code