The bit where a 40 ms hand-off gets stretched to a couple of seconds but still shows "real time: 40 ms" is a clever compromise. I've lost hours to agent traces where a silent model fallback looked identical to a normal run, so the chip on the name tag when an agent drops to another model is the feature I'd copy first. Do you also flag when the fallback model's answers start drifting, or only that the swap happened?