Ward hierarchical pooling on BEIR — 12 models, 7 datasets, fixed (maxclust) + adaptive (distance threshold) pooling