I noticed that LuaJIT has the LJ_DUALNUM define, which actually internally adds an int32 subtype which causes numbers which can fit in an int32 to be stored as such. It does not change the semantics, it's to avoid floating point ops on (mainly ARM) processors with slow or emulated FPUs. However, I think it should be possible to extend it without a huge amount of work to implement Lua 5.3 semantics with a 32 bit (rather than the default 64 bit) integer subtype. I think it would still be useful to a lot of people, personally I would love to have it.

Increasing TValues to 128 bits would be very undesirable, but since luajit already supports 64 bit ints through the ffi library I don't see why they couldn't be stored as boxed udatas/cdatas in the same way.