Fix #1045: seed and stratify the GBM propensity early-stopping split

[Hari20032005]

Fixes #1045.

GradientBoostedPropensityModel.fit() drew its early-stopping validation set with neither a random_state nor stratify, so a model constructed with an explicit seed still produced a different fit on every call.

The change

One call, in causalml/propensity.py:

X_train, X_val, y_train, y_val = train_test_split(
    X,
    y,
    test_size=stop_val_size,
    random_state=self.model_kwargs.get("random_state", 42),
    stratify=y,
)

random_state — _model resolves the XGBClassifier seed as self.model_kwargs.get("random_state", 42); reusing that expression puts the split and the classifier on the same seed. The R-learner's own early-stopping split already does this (rlearner.py:830), so this brings the last unseeded split in the library into line. random_state=None stays genuinely random, per the scikit-learn convention.

stratify — y here is the binary treatment indicator, and every other treatment split in the library stratifies on it (compute_r_residuals, whose docstring promises "every fold retains both arms"; LogisticRegressionPropensityModel's StratifiedKFold). Without it the validation set's treated count swings by an order of magnitude on imbalanced data, and at small n can reach zero, which makes the early-stopping metric meaningless.

Effect

Same seed, same data, three consecutive fits:

	before	after
early_stop=True	not identical, max abs diff 0.5466	identical
early_stop=False	identical	identical

early_stop=False was already reproducible and is unchanged — it is the control that isolates the split as the source of the nondeterminism.

Tests

Two regression tests in tests/test_propensity.py:

• ..._earlystopping_reproducible — same seed gives bit-identical scores, and a different seed still gives different scores, so the test cannot be satisfied by hard-coding a constant.
• ..._earlystopping_stratified — on 10%-treated data, asserts the validation split preserves the treatment rate.

Both fail on master and pass with the change. Verified the fix under numpy, pandas and polars inputs, which all produce identical scores.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[CLAassistant]

All committers have signed the CLA.

[Hari20032005]

The Read the Docs check failed on an intersphinx network timeout, not on anything in this diff:

WARNING: failed to reach any of the inventories with the following issues:
intersphinx inventory 'https://docs.scipy.org/doc/scipy/objects.inv' not fetchable due to
<class 'requests.exceptions.ConnectTimeout'>: ... Connection to docs.scipy.org timed out. (connect timeout=30)

Since the docs build runs under -W, that single transient warning fails it. This PR touches only causalml/propensity.py and tests/test_propensity.py — no docs, no docstrings — so a re-run should be green.

The build workflows appear to be waiting on maintainer approval (first-time contributor). Locally, tests/test_propensity.py is 9 passed on Python 3.13 / numpy 2.5.2 / scikit-learn 1.9.0.