Simulate failover with a cold cache

The simulator charges each step what the provider would: 4,000 fresh input tokens, plus the 56,000-token prefix the first time a session steps on a platform, because that step writes the cache. The fallback is a token bucket that refills at 2,000,000 tokens a minute; a step that doesn't fit gets a 429 and a retry-after. The router decides who moves each second, and the five strategies differ only in that decision. The decision that matters most is what limits the move rate: a number of sessions per minute, or the tokens the fallback can actually absorb.