Splitting from issue540, msg2135.

This needs a regression test.

Fixing this may be tricky.  

1. The hunk/binary patches should also not be removed from working
2. You'd also want to make sure that you can unpull a file add patch.

I just got bitten by this.

I would argue that revert shouldn't allow reverting of newly-added files.
Perhaps it should complain, and suggest to the user that they use `darcs
remove':

    > darcs rev file_not_recorded
    
    Error: The file `file_not_recorded' has not been recorded in the
    repository. If you want darcs to ignore the file, but *not* delete it in
    the working directory, use `darcs remove file_not_recorded', otherwise
    simply delete the file.


Darcs revert is for using darcs to undo changes that the user has made in
working, so revert is behaving as per its remit. I think that in this case,
its unlikely that a user would ever want or expect this behaviour...

I've just pushed a trivial failing test for this issue.

Eek. I think the best solution is to split revert into two phases: 
first revert pending, then calculate the diff to working, then revert 
that.