Retrieval-augmented generation (RAG) systems treat documents as flat chunk collections, discarding structural hierarchy, cross-document relationships, and causal dependencies. We present DENDRA, a multigraph document intelligence engine that constructs a 5-layer graph (structural tree, entity network, proposition facts, hierarchical summaries, and cross-document edges) as an intelligence layer over standard chunk retrieval. DENDRA introduces deterministic impact analysis — given any node modification, the system computes exact blast radius through typed dependency traversal, without LLM calls. A 14-step pluggable retrieval pipeline with adaptive routing scores chunks through graph-augmented signals. On DENDRA-Bench (400 questions, 6 domains, Russian language), DENDRA achieves 95% retrieval accuracy — +23% over plain vector RAG (72%) and +30% over LlamaIndex (65%). On legal documents: 99% vs 58% (+41%). The system runs fully offline with zero cloud dependencies.