Int8 matched float32 Recall@5 on our 100-chunk cart FAQ — ubinary needed rescore

We quantized all-MiniLM-L6-v2 on a fixed 100-chunk ScalarCart FAQ + 15 graded queries (same float32 baseline). No HF MTEB copy-paste. Desk API spend: $0. precision index bytes vs f32 P@5 Recall@5 m