StaticBuildBuild the TrainingEngine's production dependency bundle from a per-call
provider + user. Trained model files and Training Snapshot files go to the one storage provider
PS_STORAGE_PROVIDER names (default "Local File System"), each recorded as an MJ: Files row; the scoring side
reads a model file through the provider its row names (StorageProviderArtifactLoader). The snapshot store and
the artifact store share one PredictiveStudioStorage, and the engine has the artifact store resolve it before
anything of the run is written, so a provider not configured on this server refuses the run before training.
The Record Changes history the as-of assembly replays is left to the engine's default: a complete-read
RunViewRecordHistoryReader over the same dataAccess. The locked holdout is always the one the bound protocol
declares (date-carved, seeded-random or rolling-origin). Snapshots are found by content hash through the same provider and user. A unit pipeline reads the Evaluation Protocol it is bound to through the
same provider and user and trains through the trainer (MJSidecarTrainer over the process-shared sidecar).
the owning provider for data access / multi-provider correctness
the acting user threaded through every entity op for isolation/audit
StaticRetrainRetrain a completed unit run from its Training Snapshot with production deps, delegating to SnapshotRetrainer.Retrain: the stored files are verified against the snapshot row, the logged request is sent again and the holdout metrics are compared.
A retraining reproduces only on the trainer the run was trained on, so a caller outside the server that trained it
passes that trainer as trainer. Without one, the production deps' trainer is used.
the MJ: ML Training Runs row to retrain
the owning provider
the acting user
Optionaltrainer: ITrainerEvaluatoroptional trainer to retrain on (defaults to the production deps' trainer)
optional retrainer override (defaults to a fresh SnapshotRetrainer)
StaticTrainTrain a model by pipeline id, delegating to TrainingEngine.trainModel
with production deps. The engine is injectable so unit tests substitute a mock
with no live DB / sidecar (mirroring the Action's createEngine seam).
A refused run throws TrainingRefusedError unchanged: the run row already records the
refusal, and the operation surfaces its message as the operation's error.
the typed training input (pipeline id + optional overrides)
the owning provider
the acting user
optional engine override (defaults to a fresh TrainingEngine)
The training delegation every in-process caller uses: one production deps builder, one train-by-pipeline entry point and one retrain-from-snapshot entry point, so the Train operation, the Train action, auto-train, the pipeline builder and the rigs train (and retrain) the same way.