fast-import on existing branches

Hi there,

I've struggled for quite some time to sort out documented, intended and actual behavior of git fast-import. Unless I'm completely mistaken, it seems to be a straightforward bug, but if that is the case, I am really surprised why nobody else has stumbled over it before:

I managed to use fast-import for a chain of commits onto a new branch into an empty repository. I managed to use fast-import to create a new branch starting from an existing parent using the 'from' command as documented.

What I failed to do is to add commits on top of an existing branch in a new fast-import stream. As it seems, the variant of using 'commit' without 'from' only works on branches that were created within the same fast-import stream!

The different approaches I tried (each with new fast-import stream on existing repo with existing branch)
* 'commit' without 'from'
-> Error: "Not updating <branch> (new tip <hash> does not contain <hash>)
And indeed looking into the repo afterwards, a new commit exists without any parent.
* 'commit' with 'from' both naming the same branch
-> Error: "Can't create a branch from itself"
The only workarounds that I could find are to either explicitly looking up the top commit on the target branch and hand that to fast-import or create a temporary branch with a different name.

Looking through the code of fast-import.c, I can indeed lookup_branch and new_branch only deal with internal data structures and the only point were read_ref is called to actually read existing branches from the repo is in update_branch to check whether the parent was set correctly. What is missing is a call to read_ref in either lookup_branch or new_branch (probably both have to be reworked in some way to handle this cleanly). From all I can see a fix should be fairly straightforward to implement, but I am really not sure whether I have the full picture on this.

(I found all of this struggling with git-p4.py which appears to contains a complex and not fully correct mechanism to determine the 'initalParent' that appears to implement just such a workaround.)

I would be grateful for any input on this issue! Greetings, Norbert