Solving a Million-Step LLM Task with Zero Errors