[xarball]

This scares me a bit because of what they're doing to essentially drop dependency information that's preserved in normal git commit order.

If you ignore commit order, individual lines have no way of really conveying what they've been tested with, nor what they will break if added or omitted, independent of commit time.

You get a lot of that stored automatically with git, by forcing people to define changes in terms of what was there before the change takes place.

You also have the ability to reposition this information with git rebase, and to safely walk it through change conflicts to make sure you affirm what is happening.

With this in mind, I think an algorithm like darcs' history looks like it would be much more suitable in text/block based editing, rather than programming. Something like tracking history in XML or following a Microsoft Word document.

[0x0]

Absolutely, even over-eager use of "git add -p" can easily create commit "snapshots" that won't actually compile.