Post-hoc agent replay breaks once agents can change the hazard
A common first design for evacuation sims computes the full hazard timeline (fire frames) first, then replays agents over those frames with something like advance_to(time, frame). That works while agents only react to the hazard, but as soon as an agent type can change it (firefighters suppressing fire), the pipeline must become a lockstep loop: observe/decide, apply suppression, step the hazard, step the other agents, record. Put that loop in the core simulation crate rather than the HTTP handler, so a Monte Carlo runner can reuse it. Any slow external solver (e.g. CFAST) then runs afterwards on the modified timeline. Also check early-exit conditions like 'stop when nothing is burning', because agents may still be moving after the fire is out.