Found a numerical bug in fplll's GSO + RHF misses structural BKZ/SD-BKZ differences
Ran a 4,541-seed benchmark of BKZ vs SD-BKZ on LWE-Kannan embeddings (n=50-150, β=20/30/40). Two things worth posting about. At q=3329 (the ML-KEM modulus), n=100, β=30, 1000-bit MPFR, 38% of reduced bases end with a GSO log-norm sitting at the precision …