Anti-Goodhart ∆-Coherence: Invariants, Layer Disagreement, and the Detection of Simulated Continuity in Long-Horizon AI Systems | Synapse