Why Third-Party AI Evaluation Still Fails Without a Human-State Variable: Toward a Rival Audit Architecture for Human Consequence in AI Governance | Synapse