Simulate the month-end rush

The simulation keeps one FIFO queue per class, and every policy is just a rule for choosing which class's head goes next, plus an optional rule for what to do with a step when it arrives. Starts are paced by a token bucket at 213 steps a minute and a concurrency cap from Little's law, so capacity is exactly the lesson's. The decision that matters most is the per-class queues: within a class the oldest step has always waited longest, so aging only ever has to rank four heads, and the full day runs in a couple of seconds per policy instead of re-sorting a heap of 8,000 items on every start.