Session 45: Phase 5d: strip legacy.setup() from _native_setup_timeseries

The TS drain is complete. _native_setup_timeseries no longer calls legacy.setup() at all. Together with the verb drains in s40-s44, every default TS workflow runs without ever touching the legacy directory. Phase 6 (deletion of pycaret/internal/pycaret_experiment/) is now unblocked.

ADDED — engine#

• ADDED — _TSContextProxy in packages/engine/pycaret/core/experiment.py. __slots__-bound class with the 14 attrs that TS containers / util helpers read off experiment.<x>: seed, gpu_param, n_jobs_param, seasonality_present, primary_sp_to_use, strictly_positive, seasonality_type, all_sps_to_use, X_train, is_multiclass, enforce_pi, enforce_exogenous, exogenous_present, fe_target_rr, index_type. fe_target_rr defaults to None to match legacy default — [] would break the cds_dt recursive forecasters via truncate_start validation.
• ADDED — Experiment._auto_detect_seasonality(y) static method. Lightweight port of legacy seasonality detection: derive candidate sp from y.index.freqstr via pycaret.utils.time_series.get_sp_from_str, run sktime's autocorrelation_seasonality_test on each, return (seasonality_present, primary_sp_to_use, all_sps_to_use) where primary is the largest significant sp.
• ADDED — Experiment._build_ts_fold_generator(...). Same math legacy uses: step_length = len(fh), initial_window = len(y_train) - ((fold - 1) * step_length + max(fh)), with defensive shrink to n_train // 2 when folds wouldn't fit. Returns ExpandingWindowSplitter (default) or SlidingWindowSplitter based on fold_strategy.

CHANGED — engine#

• CHANGED — Experiment._native_setup_timeseries rewritten end-to-end. No legacy.setup() call. Coerces input → univariate Series + optional exogenous DataFrame, builds ForecastingHorizon from self.fh, auto-detects seasonality, splits via temporal_train_test_split, builds fold generator + model registry through the proxy, and constructs a minimal ForecastingPipeline (placeholder NaiveForecaster) as preprocess_pipeline.
• CHANGED — Experiment.models() now filters by model_type ∈ TSModelTypes directly on the snapshot registry for TS, matching legacy's filter that excludes ensemble_forecaster. No more legacy.models() deferral for TS.
• CHANGED — TimeSeriesExperiment.predict_model detects fitted bare forecasters (via is_fitted attr or presence of _y) and uses them directly instead of wrapping in the placeholder pipeline (which is unfit and would invalidate the forecaster's state).

ADDED — tests#

• ADDED — packages/engine/tests/test_session45_ts_native_setup_full_drain.py — 9 new tests:
  • Drain-lock for legacy.setup (poison + verify native succeeds with _native_setup_used is True).
  • Seasonality auto-detection on monthly airline → seasonality_present=True, sp=12, seasonality_type ∈ {mul, add}, strictly_positive=True.
  • No-frequency fallback → (False, 1, [1]).
  • ExpandingWindowSplitter default with n_splits == self.fold.
  • SlidingWindowSplitter when fold_strategy='sliding'.
  • Full sktime model registry populates through the proxy (naive / arima / ets / theta / etc. all present).
  • End-to-end create + tune + finalize + predict chain with all 6 legacy verbs poisoned including setup.
  • setup_kwargs still routes to legacy (compatibility preserved).
  • Univariate Series input accepted directly.

INTERNAL#

• INTERNAL — Why a proxy instead of a real legacy instance. The proxy is a 30-line __slots__ class. A real legacy instance pulls in 10K LoC and runs Fourier analysis + a dozen other pre-flight checks before even getting to model construction. The proxy is fast (<1ms to instantiate) and has zero dependency on the directory we're about to delete in phase 6.
• INTERNAL — fe_target_rr=None vs []. Legacy initializes fe_target_rr to None (unset) by default. Containers either pass it through unchanged (cds_dt recursive forecasters expect None so they use their internal default), or replace it. Passing [] instead causes the cds_dt models to try to apply zero feature engineering steps and trips sktime's Reduce.fit validation: "Reduce must either have window length as argument or needs to have it passed by transformer via truncate_start". One-character fix; significant test recovery.
• INTERNAL — Why Experiment.models() no longer defers to legacy for TS. With phase 5d, _fit_state["model_registry"] is fully populated through the proxy — it has the same _all_models_internal shape as legacy. The legacy filter (model_type in TSModelTypes) is applied directly on the snapshot. No reason to call legacy.models() anymore, and doing so would crash anyway because legacy.setup() was never called → no legacy.seed.
• INTERNAL — Why bare-forecaster bypass in predict_model. The placeholder ForecastingPipeline we create in _native_setup_timeseries is unfit (it's a template). When the user passes an already-fitted bare forecaster to predict_model, wrapping it via _add_model_to_pipeline produces an unfit pipeline (the wrapper fits all steps top-down), so pipeline.predict(fh) raises NotFittedError. The fix: detect that the bare forecaster is already fitted (via is_fitted or _y) and use it directly.

Session 45 delta summary#