nirs4all.api.general_retrain module

Resolve a captured winner into fresh estimators for a full DAG retrain.

Sources follow the existing trusted-host artifact integrity policy. Cloning their constructor parameters must not retain estimator objects, rerun a generator search, or make these Python artifacts portable Core archives.

nirs4all.api.general_retrain.captured_training_spec(source: Any, *, mode: str = 'full', new_model: Any = None) → tuple[list[Any], dict[str, Any]] | None[source]

Select a recorded host winner before execution, without a retry fallback.

Existing bundles with a training specification retain their published replay behavior. Prediction dictionaries resolve only recorded workspace chains; caller-provided model objects or parameters are not authoritative.