Latency budgets forced us to drop nice-to-haves. Two cents after running this through prod traffic. Optimize observability before model tuning.