We're not (yet) trying to optimize the code being executed (we should perhaps add an issue about class creation to https://github.com/faster-cpython/ideas).

I agree that marshal is very fast already; I doubt the extra cost of wordcode would pay off since it would increase the size. This could be tested but it would be a big experiment.

I'm not sure that compression will really help, since a typical decompressor delivers its output in some memory buffer. But this needs to be tested, and it might be a simpler experiment.

I'm pretty sure zipimport will improve things -- the main problem with it (as always) is that not all tooling understands it, since resource loading is different.