To be honest, their insistence to port this to Rust is likely the culprit of the performance issues.

Let's say there was a 1:1 decomp instead, bit-by-bit exact. This is doable today - what then?