You switched to binary quantization and recall@10 fell eight points. Diagnose it.

Lead with the mechanism: Hamming distance estimates θ/π with binomial variance. At d = 768, this variance is roughly ±0.027 on cosine, meaning any pair of candidates closer than about 0.04 is now decided by noise. You should then check the three causes in order: no rescoring stage, no rotation before the sign, and a corpus whose gap distribution is genuinely too tight for one bit. These should be separated by measuring the distribution on queries for which you already have labels.