Scaling Pain of Coding Agent Serving: Lessons from Debugging GLM-5 at Scale