Replay stored decisions under a candidate policy

Retro-simulation: POST a partial candidate policy (dollars, same shape as the policy update) and a range up to 92 days (defaults to the last 7); every stored decision in the window re-runs through the pure ladders with the candidate overlaid on its recorded context, and the response reports what flips — totals by effect, approved-spend delta, and the changed decisions with before/after codes. Read + compute only: nothing is persisted, debited, or escalated. mode=as_recorded (SIM-1, default) holds budget counters as the engine saw them; mode=sequential (SIM-2) replays in chronological order threading each agent's approved spend forward, so an early would-denial frees budget for a later request. Owner-only.