Not quite 100% done analyzing the data, but here is what I think I'll do:
- respect the 3 plane convention. Unicode has 17 planes, with the first 3 in active use. Plane 14 is used, but it is specific and only has 368 allocated code points. It is so specific that I'll add exclude it, the same as I do all planes except planes 0-2. All excluded planes have the property that any function f(cp) = cp.
- I'll ignore all special substitutions: sub and superscript, font, circle, square, fractions and of course compatibility substitutions. I won't be using a flag for now, just exclude them.
- I'll ignore all CJK COMPATIBILITY IDEOGRAPHs. There is no way a general purpose library can provide satisfactory use case for these. If you really needs such substitution, you will probably use a more competent third party library. f(CJK COMPATIBILITY IDEOGRAPH) = CJK COMPATIBILITY IDEOGRAPH

All these combined with my two table solution, with a chunk size of 256 to 1024 will leave me with around 8000-9000 bytes of data in each executable that does decomposition. Final numbers will be determined once implementation is done and round trip testing is complete.

I think this is a reasonable subset that can handle NFD, at the small price of a flat 9K in exe size, + the size of the actual methods.