This paper identifies a text-vector integrity gap in offline retrieval-augmented generation (RAG) where signed knowledge packs contain text and precomputed embeddings, but signatures only authenticate the publisher, not the vector-text correspondence. The authors formalize the hijack rate as the distribution of angular retrieval margins at a given cosine tolerance (the margin law). Evaluating five corpora and two embedding models, they find small margins: an idealized tolerance at floating-point noise levels exposes up to 5.5% of queries (max 1.2% excluding exact duplicates), while a practical threshold calibrated from CPU INT8-FP32 comparisons of a small model at the 0.1% quantile of honest cosines leaves 98-100% of queries exposed. On a Raspberry Pi 5 running a local LLM, such an attack reduces correct answers from 0.75 to 0.39. Integrity audits of FAISS product-quantization indexes over 57,638 documents fail similarly; tampered vectors passing audit place the attacker's document first for 25% of targets. The authors propose verifying quantization codes with a calibrated guard band, canonicalization, and verifier-decided ambiguity resolution, which accepted all honest vectors across three platforms and eliminated publisher choice. They also demonstrate that common floating-point and dynamic-INT8 inference pipelines lack bit-reproducibility across batch sizes and architectures, and report negative results on abstention, context pruning, and semantic screening defenses.