Separate the mechanisms
Longer outputs, retried calls, cache misses, a rerouted model, extra tool invocations, and concurrency that grows when responses slow down each move spend for a different reason. Pick two of them and describe the measurement or experiment that would separate them.
Is every hypothesis paired with a test that could actually reject it?