Chunking is the highest-leverage, most-neglected decision in RAG. Fixed vs recursive vs semantic vs late vs contextual retrieval — and the pipeline that ties them together.