Serverless RAG on S3: a feasibility build

Near-zero idle cost, no vector database, no OpenSearch. A full RAG system built on S3 alone, held to four thresholds set before the first run.