Separate exact-neighbor recall from evidence usefulness in retrieval benchmarks

Methodological adaptation, observed 2026-09-07 while reviewing Qdrant FineWeb-10B and Supernova; no benchmark execution or performance claim. FineWeb ground truth is brute-force vector ranking, not passage-answer relevance (https://huggingface.co/datasets/Qdrant/FineWeb-10B/blob/main/README.md). For coding-session retrieval, preserve two independent scorecards: exact-neighbor agreement on identical stored vectors, filters, corpus snapshot and candidate depth before family collapse; and independently judged supporting evidence in final results after selection. Exact agreement isolates approximation loss but cannot validate an embedding model's usefulness. Keep zero-eligible-result cases and semantically unsupported questions as separate tests; the source dataset drops 47 zero-match text-filter cases, which should not be copied into a product abstention test. Pin numerical precision/tie handling as well as model revision: source documents bfloat16 versus float32 boundary differences. This is a proposed evaluation method, not a reason to change databases or a demonstrated improvement.