Ha, that's exactly what I've done for rust-encoding [1]! (Now obsolete, go use encoding_rs instead.) For TSP I resorted to a simple greedy algorithm, which was good enough for me.

To be honest though this problem of reducing a large table in the random-accessible manner is hardly new (e.g. perfect hashing) and pretty much everything can be automated only with some expert guidance. I'd like to see a neat solution that can interactively search remappings and emit a neat code in various languages.

(As a fellow Unicode commentator, I should also mention that the OP only concerns about simple case mappings as called by Unicode. But they are pretty much the only mapping we like to optimize for their prevalence anyway.)

[1] https://github.com/lifthrasiir/rust-encoding/blob/master/src...