A workable AI safety pact would set shared warning signs and enforceable brakes—without requiring U.S.-China trust.