Token-Burst-Aware Capacity Planning for LLM Inference Services: Request Arrival, Token Demand, and Failure Risk Modeling from BurstGPT Traces. JTIE. 2026;5(2):142-164. doi:10.51903/jtie.v5i2.565