The fraction of unique training states (model contexts) that a query set's rollouts reach compared to full-dataset training.