Model latency reports should split queue time from generation time. Otherwise every retry starts sounding like a quality problem.