Session 30: Internal-state drain: transformed splits + fold generator + model registry

Baseline: session 29 drained the 7 user-facing data accessors. Session 30 drains the 6 internal self._legacy.<x> reads still present inside the drained verbs.

CHANGED — engine#

• CHANGED — packages/engine/pycaret/core/experiment.py — _snapshot_fit_state() captures 6 new slots in _fit_state at the end of fit(): X_transformed, X_train_transformed, y_transformed, y_train_transformed, fold_generator, model_registry (a dict(legacy._all_models_internal) copy).
• CHANGED — 13 callsites drained across experiment.py, supervised.py, unsupervised.py:
  • create_model (supervised native) — reads X_train_transformed, y_train_transformed, fold_generator, and model_registry from _fit_state.
  • _resolve_supervised_estimator — reads model_registry.
  • tune_model — X_train_transformed, y_train_transformed, fold_generator, model_registry.
  • compare_models — model_registry.
  • stack_models — fold_generator.
  • calibrate_model — fold_generator.
  • finalize_model — X_transformed, y_transformed.
  • Unsupervised create_model + assign_model — X_transformed, model_registry.

ADDED — tests#

• ADDED — packages/engine/tests/test_session30_internal_state_drain.py — 5 tests with a generalised drain-lock helper (_PoisonedAttrAccess) that raises on every dunder a verb might use to read legacy state (__getattr__, __getitem__, __contains__, __iter__, __len__). Tests cover:
  • create_model succeeds after every internal-state attr on _legacy is poisoned.
  • tune_model resolves the search space from the snapshot's model_registry even when legacy._all_models_internal is dropped to {}.
  • finalize_model re-fits using _fit_state["X_transformed"] rather than legacy.
  • Clustering create_model fits on the snapshot's X_transformed.
  • Sanity: all 13 _fit_state keys are populated post-fit.

INTERNAL#

• INTERNAL — Why a copy of the registry, not just a reference. _all_models_internal is mutable on the legacy class — extensions like add_metric could grow it. Taking a dict(legacy._all_models_internal) shallow-copy at fit-time makes the snapshot stable: future legacy mutations don't leak into already-fitted experiments. Trade-off: if a user genuinely wants their add_metric call to affect a previously-fitted experiment, they'd need to re-snapshot. That's an edge case worth the reproducibility win.
• INTERNAL — _PoisonedAttrAccess is the drain-lock pattern at scale. Sessions 22-28 monkeypatched a single legacy method to raise; session 29 monkeypatched a property; session 30 needs to lock down ~6 attrs across multiple dunder methods. Wrapping them all in one sentinel object that raises on every access keeps the test code short and the failure mode clear ("a verb reached for self._legacy.<x>").

Session 30 delta summary#