Within-compound associations explain potentiation and failure to overshadow learning based on geometry by discrete landmarks.