Add ServiceMonitor for llm-serving namespace. Wire prometheus.io annotations to all vLLM pods (reasoning, ornith, embeddings, reranker). Scrape /metrics@8080 every 30s with proper relabeling.