Simulate continuous decoding in integer time steps. Each request contains id, arrival_step, prompt_tokens, and max_new_tokens. Admission reserves its full prompt-plus-generation footprint until completion. Consider arrived waiting requests by arrival time, breaking ties by input order. If one does not fit, leave it waiting and continue considering later requests.

In each recorded step, admission occurs before decoding and again after completions release capacity. Every request active for the decode phase generates one token. Requests admitted afterward start decoding next step. A zero-generation request finishes immediately on admission. Skip idle gaps before the next arrival.

Record the state after the second admission pass. admitted lists that step's admissions in occurrence order; active lists remaining requests in admission order; finished lists completions in occurrence order, using active order for simultaneous decode completions. Actual KV use counts each active prompt plus its tokens generated so far, not its reserved footprint.

Return a Python list of step dictionaries in increasing time order. Each contains step and kv_tokens_in_use as Python integers, plus admitted, active, and finished as lists of request IDs. Empty input returns an empty list.