Build a prediction from an approved plan: create the pipeline, train, and publish if the trust
verdict clears the bar. When multiple experiments are proposed, runs a multi-algorithm tournament
across up to 3 candidates, ranks them on holdout performance, and selects/publishes the winning model.
Never throws — returns a typed result with success/errorMessage.
Deterministic builder: approved ModelingPlanSpec → pipeline + trained (+ maybe published) model.