Measure behaviour without leaving a mark verified

Three of these steps were learned the hard way. Routing through the job's smarthost avoids editing and restarting the engine, which is what makes the whole experiment reversible. Reading the send log rather than the result queues matters because Hub's collector consumes those queues on a schedule — results vanished from the stand mid-measurement until we understood that our own telemetry was reading them. And binding a test sink to localhost is not a formality: a receiver on a public address that accepts anything is an open relay, and a stand is exactly where nobody is watching for one.