This paper presents ARC M4.0, a dual-reasoning engine architecture designed for long-horizon operation of humanoid and embodied AI systems. Building on prior ARC research, this work formalizes a strict separation between interpretation authority and execution authority at the operating system level. Through empirical stability experiments, failure-pressure simulations, and long-term session analysis, ARC M4.0 demonstrates that structural decoupling of reasoning and execution prevents authority escalation, identity drift, and irreversible failure during physical operation. The results indicate that stable autonomy in embodied AI is achievable not through additional constraints or rules, but through fixed authority boundaries embedded in system architecture. ARC M4.0 positions dual-reasoning as a foundational design principle for future humanoid AI operating systems.