Prompt/decode rates are derived from the server counters (Δtokens / Δbusy-time) and attributed across the busy window — the server only flushes some counters when a task completes.