Replication package: diagnostic evaluation of text embeddings on multi-dimensional similarity (code and derived results, v7)

Anonymous replication materials: the full analysis pipeline and every derived quantitative output cited in the associated article, spanning three benchmark branches - canonical long-form fiction, newswire text, and open-access scientific articles - with the chunk-position contamination probe and its position-matched control, the second large-family (multilingual retriever) training-regime branch, mixed-effects convergence checks, deep-bootstrap robustness checks, a closed-form derivation check for the view-fusion weights, and an independent audit script that recomputes every reported number (152 claims) from the released derived data. v7 refreshes the archive alongside the final camera-ready revision; scientific content is unchanged from v6 (no new experiments, no new numbers). Code: MIT. Derived data and annotations: CC-BY-4.0. No personal data are involved.