Measure active CPU against wall time

The app is one endpoint written three ways, so the only thing that changes between load tests is how the handler waits. A second FastAPI app stands in for the model and only sleeps, so nothing paid is called. The decision that matters most is how CPU is measured: time.process_time() is per process, not per request, so under concurrency it counts every request that ran while yours was waiting. The answer measures CPU three ways and the difference between them is one of the findings.