One thing I’d really like to build, if I ever have the time/energy, is a static analysis tool that automatically determines an optimal subset based on actual usage. Subsetting arbitrarily based on say charset definitely helps, but it can go even further if you only use even a subset of that charset.