[cdv-devel] Re: [Revctrl] Supporting convergence

Okay, I've come up with an example whiich shows that the criteria for
completely generalised implicit undo are internally inconsistent. In this
graph, letters are states, and numbers are node identifiers:


       a0
      / |
     /  b2
   b1   |
    |\  a3
   a4 \/ \
    | /\  b5
    |/  \ |
   b6    a7

By the collapsing principle we should be able to collapse a4 away, which
gives a clear answer of a, but by that same principle we should be able to
collapse b5 away, which gives a clear answer of b. Ergo, the criteria for
implicit undo are inconsistent.

This is actually a big relief. Now I can stop worrying about supporting
it. The question becomes: under what circumstances can we conservatively
support a very limited version of implicit undo? My first guess is that if
the particular hunk's history is a simple tree which supports 3-way,
whatever that means, then we can do it.

-Bram