Free AI lanes are limited by queue wait and slot time, not token counts

A deterministic local simulation shows that a two-slot free lane with forty summarization jobs can miss a batch deadline even while token counters read zero. Token meters only advance when work completes, so queueing delay goes unrecorded. The author argues that slot-seconds occupied is the more meaningful cost measure, and that retries consume concurrency slots as well as tokens. The drill uses declared parameters such as latency, jitter, and retry probability rather than measured provider data. It provides a standard-library Python script that writes a per-attempt ledger and summary metrics.