24x speedup (over 10x already) and similar loss profile (for c++ version, optimized by claude): https://gist.github.com/freakynit/3982eab8413a89941bd0018e63......