[John Clow (Gamrix)]

These are three gaps in stubs/sklearn that produce false type errors on current scikit-learn code. They affect Pylance users and also Pyrefly users, because Pyrefly bundles these stubs. I found them by running Pyrefly against a codebase on scikit-learn 1.9.0. Each signature below was checked against both 1.9.0 and the 1.6.1 runtime that CI's stubtest installs (scikit-learn <1.7.0).

1. sklearn.cluster.HDBSCAN is missing (added in 1.3)

from sklearn.cluster import HDBSCAN fails because the stub doesn't export it.

• Adds cluster/_hdbscan/__init__.pyi and cluster/_hdbscan/hdbscan.pyi, written in the style of _dbscan.pyi and _optics.pyi: constructor parameters, fit, fit_predict, dbscan_clustering, and the fitted attributes labels_, probabilities_, centroids_, medoids_, n_features_in_ and feature_names_in_.
• Exports HDBSCAN from cluster/__init__.pyi, including __all__.
• Source: sklearn/cluster/_hdbscan/hdbscan.py#L423 and __init__ at L665.
• copy is typed bool = False. In 1.9 the runtime default is a hidden "warn" sentinel (Hidden(StrOptions({"warn"}))) that behaves like False and raises a FutureWarning. In 1.6 the default is plain False, which stubtest checks.

2. sklearn.tree._tree.Tree lacks its public node-array properties

estimator.tree_.children_left, .feature, .threshold and the other node arrays are the documented way to inspect a fitted tree, but none of them are in the stub.

• Adds read-only properties children_left, children_right, feature, threshold, impurity, n_node_samples, weighted_n_node_samples, missing_go_to_left (all ndarray) and n_leaves (int, which matches BaseDecisionTree.get_n_leaves() -> int). Source: sklearn/tree/_tree.pyx#L721-L761.
• Fixes the typo max_dept: int to max_depth: int (cdef public intp_t max_depth, _tree.pxd#L50).
• Changes n_classes from int to an ndarray property. At runtime it is a property that returns an array with one entry per output (_tree.pyx#L721-L723).

3. precision_recall_curve lacks drop_intermediate and y_score

• drop_intermediate: bool = False was added in 1.3.
• The second parameter was renamed probas_pred to y_score in 1.5, with probas_pred deprecated and removed in 1.7. The 1.9 signature is (y_true, y_score, *, pos_label=None, sample_weight=None, drop_intermediate=False) (sklearn/metrics/_ranking.py#L1059).
• This follows the existing squared deprecation pattern in metrics/_regression.pyi. The main overload takes y_score. A second overload, marked @deprecated, accepts the keyword probas_pred= that 1.5 and 1.6 still allow. Positional calls precision_recall_curve(y, scores) are unaffected.

stubtest_allowlist.txt

• Removes the entries these changes make unused: sklearn.cluster.HDBSCAN, sklearn.cluster.__all__, sklearn.cluster._hdbscan.hdbscan, the ten sklearn.tree._tree.Tree.* properties and max_depth, and both precision_recall_curve entries.
• Adds entries for internals the new module doesn't stub, in the same way sibling estimators handle them: HDBSCAN.__sklearn_tags__ (both paths), hdbscan.FAST_METRICS and hdbscan.remap_single_linkage_tree.

Validation (local, Python 3.13, --group tests)

• python tests/run_stubtest.py sklearn reports exactly the same set of errors before and after this change, so the PR adds no new stubtest errors and leaves no unused allowlist entries. Two caveats:
  • I had to pin numpy<2.3 locally. With numpy 2.5 the run stops with an unrelated build error: sklearn/utils/validation.pyi: Module "numpy.core.numeric" has no attribute "ComplexWarning".
  • Even pinned, the run reports 63 errors on main that this PR doesn't touch (_seq_dataset disjoint-base errors and others), so a CI run may still fail on those.
• python -m mypy .: no issues.
• pyright stubs/sklearn tests: 0 errors. A probe file checking HDBSCAN(...).fit(X).labels_, tree_.children_left, tree_.n_leaves and precision_recall_curve(..., drop_intermediate=True) also checks cleanly.
• ruff check stubs/sklearn and ruff format --check stubs/sklearn pass. Repo-wide ruff check reports one existing RUF100 in stubs/matplotlib/pyplot.pyi, which this PR doesn't touch.

🤖 Generated with Claude Code

[John Clow (Gamrix)]

@microsoft-github-policy-service agree company="Safetykit"