Turns graded examples and traces into a training set: selects which rows earn a place, holds evaluation items out by identifier at the one seam where that is still possible, shapes rows to the target method, masks loss to response spans, and emits a provenance card that gates the run. Use when building or auditing training data.
