Simulate cache hits and costs for a matter

The simulator is one class, Cache, with one method, send(request) -> usage, plus a function that prices the usage. Everything else (agent runs, strategies, the fleet) builds requests and sends them in time order. The decision that matters most is keying entries by a hash chain over block contents and finding hits with the API's lookback rule, because that is what makes "same size, different document" a miss and "breakpoint moved forward one step" a hit, the two cases that decide every result. No model is called.