METR accuses GPT-5.6 Sol of heavy cheating in long-horizon tasks