v0.4.2

Added

• TabPFNEmbedding now provides a standard scikit-learn fit / transform / fit_transform interface, so it composes with sklearn workflows. fit_transform produces out-of-fold embeddings when n_fold >= 2; transform uses a final full-data model. The previous get_embeddings(...) method is retained for backward compatibility. (#298)
• interpretability.shapiq_to_shap_explanation(...) bridge helper (new interpretability/shap.py module) that converts a shapiq explanation into a shap.Explanation, so you can compute Shapley values with shapiq but plot with the shap library without writing the loop/stack/baseline boilerplate. (#292)
• New examples: bar-distribution / predictive-distribution plotting for regression uncertainty (examples/predictive_distribution/) (#299), and multi-output regression + multi-label classification via sklearn's MultiOutputRegressor / MultiOutputClassifier (examples/multioutput/) (#297).
• Add an automated release pipeline based on TabPFN's pattern: workflows for opening release PRs, auto-tagging on merge, and publishing to PyPI via trusted publishing with a manual approval gate. Per-PR changelog fragments via Towncrier (see changelog/README.md). (#278)

Changed

• feature_selection now returns a structured FeatureSelectionResult (carrying the selected features and the before/after cross-validation scores) instead of the raw sequential-feature-selector object, and logs progress when verbose is set. CV scores are now always computed and returned. (#293)
• TabPFNEmbedding.fit(...) now returns self (was None) and takes X, y (was X_train, y_train), to follow the scikit-learn estimator protocol. (#298)

Fixed

• Numerical overflow in the unsupervised outlier-detection "diffuse density" combiner. outliers() now computes the arithmetic mean of densities in log space via the logsumexp identity, fixing overflow on diffuse-density inputs. (#291, closes #289)

Removed

• The empty classifier_as_regressor install extra and its stale README reference. The underlying module was removed long ago; the extra was a no-op. (#300)