A One-Page Ledger Method for Generating Voynich-Like Text

[ReneZ]

Let alone why a great majority of possible modifications are never made.

No, the manuscript never explored every possible word combination. A realistic copy-mutate system would stay conservative, reusing active word families and nearby variants instead of wandering randomly through all legal forms.  

Why would a realistic copy-mutate system stay conservative? There is nothing that dictates this. 
The number of rules observed over tens of thousands of words are quite complex, and they are indeed rules.

This is, in a way, backwards logic.
We see that the word variations are very strict. Therefore, if the text was generated by modifying previous words, it would have to have followed strict rules. That is the correct direction of the logic.
There is no reason to assume that there would be very strict rules (which are then broken somewhat gradually).

EDIT:
Let's do some rough counts.
The word chedy could be considered to have four characters. 
Limiting to edit distance 1:
Each of these could be changed into another, leading to 4 times, say, 20 options.
Each of these could be deleted, leading to 4 more.
A new character could be added in each of 5 slots, so 5 times 20 more.
6 pairs could be swapped (not sure if that counts as edit distance 1).
We are close to 200 alternatives. 
Possibly 10 exist.

We can consider two alternative methods for a creation of a meaningless text using word permutations.

Method A: 
first, a vocabulary is set up using word patterns and their variations
then, a text is composed by somewhat aribitrarily picking words from this vocabulary/dictionary

Method B:
a text is generated by creating new words from previous ones 'on the fly'
then, the resulting vocabulary is the collection of all these words

It should be clear that the very limited set of allowed permutations much better fits with method A than method B.