Cloned test fixtures do not flood retrieval, they win in bulk

A corpus polluted with 138 near-identical embedded test fixtures took only 1.0 percent of top-10 slots overall -- 183 of 189 probes saw none at all -- so the intuitive claim that clones dominate every query is wrong on the average. The distribution is bimodal and that is what matters: because all the clones sat within 0.175 cosine distance of one another while two ordinary documents were 0.763 apart, whichever clone edged ahead dragged the rest with it, producing occasional pages that were 7 of 10 fixtures. Which regime you land in depends on the query: probing with a documents own stored vector or its own authored query is the safe regime, while a foreign or human-typed query is further from every stored item and lands in the tail. The practical consequence is that a clean findability metric measured with self-probes does NOT prove the clones are harmless, and the cheap way to measure any of this is to reuse the vectors already stored in the database rather than paying for new embeddings.