Session 33: get_config / set_config drain

Baseline: session 32 drained add_metric / remove_metric. Session 33 drains the last two drainable secondary verbs — get_config / set_config. Every drainable verb on the public surface is now native; only setup() inside fit() and the Phase-3 plot verbs remain on _legacy.

CHANGED — engine#

• CHANGED — Experiment.get_config(variable=None) drained. Reads from self._fit_state (data accessors, transformed splits, registries) + constructor params on self. With variable=None returns the sorted list of accessible names. Raises ValueError for unknown names. Includes legacy aliases: seed → self.session_id, pipeline → _fit_state["preprocess_pipeline"].
• CHANGED, BREAKING — Experiment.set_config(variable=None, value=None, **kwargs) drained, with a tight allowlist. _SETTABLE_CONFIG_KEYS = {"session_id", "n_jobs", "verbose", "fold", "log_experiment"}. Anything outside this set raises ValueError (with a message pointing to re-creating the Experiment). Underscore-prefixed names are explicitly rejected. Mixing positional variable + **kwargs raises (legacy semantics).
• CHANGED, BREAKING — set_config no longer silently mutates snapshot-invalidating params. Legacy set_config("target", "newcol") would silently update self.target but NOT propagate to the snapshot, leading to a confusing partial-mutation state. The new version refuses + tells the user.

ADDED — tests#

• ADDED — packages/engine/tests/test_session33_config.py — 10 tests covering: full-list get_config(None), snapshot identity reads (get_config("X_train") is exp.X_train), seed / pipeline aliases, unknown-variable rejection, single + bulk set_config, non-settable / underscore / mixed-call-shape rejection. Drain-locks for both verbs.

INTERNAL#

• INTERNAL — Tight allowlist rationale. Legacy set_config allowed mutation of any _variable_keys.difference(_property_keys) entry, which was a moving target. Tight allowlist makes the contract explicit + future-proof: adding a new safe-to-mutate param is one line in _SETTABLE_CONFIG_KEYS. Mutating something that affects the snapshot (target, train_size, fold_strategy) is properly an error in 4.0 because the snapshot won't update.
• INTERNAL — Why aliases. seed is more idiomatic than session_id (sklearn / numpy convention), but the constructor takes session_id for clarity (it doesn't seed a global state — it's the per-Experiment session ID that becomes the random_state for downstream models). Supporting both names in get_config matches what notebook code already uses.

Session 33 delta summary#