Turn preference feedback into a concrete training config.
Recent revisions are kept. Restoring one creates a new revision.
