“Would you bet 100$ on this new theory, Prof. Feynman?” — “No, but 10$ I would.”
(Richard Feynman, attributed)
This is my personal pet theory which I’ve entertained for some time.
I’m not really convinced that I’ve got the correct solution here, but I think this scheme would explain several of the characteristics of the VM, while not being completely implausible. Unfortunately, this theory introduces a whole lot of degrees of freedom, so it’s a bit difficult to actually test.
Anyway, I’ll explain the general idea and let you decide for yourselves. The basic idea is to disassemble the letters which make up the plaintext into their individual penstrokes (or, if you start with printed text, disassemble them into their graphical elements). Later, each of these penstrokes is assigned one of the VM letters — voila, instant encipherment!
See a detailled discussion of the Mark I of this algorithm. I wasn’t really statisfied with it, but when I came across Robert Firths notes, I had a relevation and modified the algorithm to Mark II, its current incarnation.
Thoughts about the Voynich Manuscript 
You are c l o s e to the “folded Key” answer !
-=se=-
Makes sense to me.
Just a weidr thought: Has anyone ever considered that the word boundaries may be shifted to one side adding extra “security”? Itw ouldm akeo nev erys implec ypher tow irtey eto neh ardert or ead.
I don’t think anybody ever looked at it from this particular angle. But considering that some phrases are almost always word-initial in the ciphertext (“qo”, for example) and other are word-terminal (“dy”), I’m quite convinced that ciphertext words do not map 1:1 to plaintext words anyway. It would also be hard to reconcile that find with your model of shifted spaces.