Real-time stream of gateway events via Server-Sent Events. Each row is a completed inference request — tenant, model, token counts, cost, and latency — emitted within 2 seconds of completion.