refactor causaliq-data modules according to module and method rules¶

All causaliq-data modules were refactored to obey the development policies.

Constraints learned from other package refactors¶

Behaviour must stay byte-identical
Preserve import binding sites
No behaviour-free bulk test rewrites: add focused tests only where coverage gaps appear

Commits¶

Stage the work as between two and four commits which the user will do:

Code and tests — one to three commits
Documentation — one documentation commit

Do not mix documentation-only changes into the code commit.

Gates after each slice (as in `scripts/check_ci.ps1`)¶

black --check src tests
isort --check-only src tests
flake8 src tests
mypy src/
python scripts/check_code_rules.py --strict src/causaliq_data

100 % coverage is maintained on the new modules and the bnlearn score replication tests in tests/functional/test_score_bnlearn.py stay green.

Outcome¶

The refactor is complete. python scripts/check_code_rules.py reports zero breaches across all 14 modules in src/causaliq_data, the full test suite passes with 100 % coverage, and behaviour is unchanged.

Stage	Modules refactored	Breaches reported
Starting point	–	19
Slice 1	`__init__`, `indep`, `pandas`	13
Slice 2	`numpy`, `numpy_counting`, `numpy_categories`	9
Slice 3	`score`, `score_categorical`, `score_gaussian`, `score_params`	0

What changed¶

Long methods and functions were split into private helpers, keeping every method within the statement and nesting depth limits.
Modules that exceeded the module statement limit were decomposed into sibling sub-modules which the original module re-exports, so public API, module paths and import sites are unchanged. Details of the decomposition are recorded in the architecture overview.
Modules whose functions are substituted in tests with monkeypatch.setattr were flattened in place (__init__, indep, pandas.read, pandas.marginals, numpy.read, numpy.from_df, numpy.set_N, numpy.marginals, numpy.unique and the whole score dispatch layer), so patches such as causaliq_data.score.gaussian_node_score, causaliq_data.score.categorical_node_score, causaliq_data.numpy.MAX_CATEGORY and causaliq_data.numpy.NumPy.MAX_BINCOUNT still intercept the calls.
Floating point accumulation order was preserved: helpers return intermediates and the caller accumulates them in the original sequence (BGE, entropy scores, Bayesian scores and the chi-squared/MI statistics).
No test was rewritten or moved. The unchanged suite still reports 100 % coverage for every new and re-structured module.

Verification¶

Every slice was checked with black --check, isort --check-only, flake8, mypy src/ and the strict rules report, followed by the full test suite at 100 % coverage on Python 3.9 to 3.13 in CI.

Stage	Statements (coverage)	Coverage
Starting point	879	100 %
Refactored	1059	100 %