Optimise later, I would say. And benchmarks in this case are going to be really weird. Too many different filesystems with different penalties, with or without caching, file or db cache, with or without opcaches, ....
Exactly. Developers (myself included!) tend to optimize way too fast/ early. I'd start with the bare minimal, get that into core, spend some time evaluating, and improve as necessary. Making the best possible case for designers is exactly what this is about, and this time around, we should go all the way.
I've done a rough split of the theme functions from the modules based on the current 4.7 (didn't have the latest cvs, but it should be similar). If you want I can prepare a patch and submit it on drupal.org together with the split code (slow, ugly, but quick to code).
I've read your blog post and am particularly interested in trying to test this as soon it is working as intended. Step #1 is to run the scripts and to do some performance benchmark. Have you done any experiments yet? When can we expect to be able to play with this? :-) Thanks for taking a lead in this, Vlado! Much appreciated. -- Dries Buytaert :: http://www.buytaert.net/