Rebased to work on top of @jdavid 's refactor/fix. The first makes sure we have a "peelable" interface for both references and objects, so we can get to objects regardless of the input.

The second one then makes use of that in order to peel whatever non-string we got, and forcing the string to be a valid revspec, which we then continue trying to peel (so we can still get to "HEAD~"'s tree). I think the current version would not work well if there are annotated tags involved.

BTW, since it's mildly related, I was also playing around with moving diff() out of Repository. There is are a couple of cases where the repo object is needed (revparse, and diffing to worktree or index). No other case (and this includes once we've run revparse_single()) needs the repository at all, as the diffs actually happen between the objects themselves (and in general there's no need for the objects to belong to the same repo, which makes the need to choose a particular repository in which to call the method rather an arbitrary restriction. We can even run diffs against arbitrary pieces of text (though it's not wrapped yet)

So I was thinking of having pygit2.diff() which takes objects and keep Repository.diff() for the cases where do want to use the index/worktree or run a revparse query, which would then fall back to the general diff, which would build on top of this.

There is already Reference.get_object() that wraps git_reference_peel, we could do with just one wrapper. And for coherence with Object.peel(...) it should be named Reference.peel(...), so Reference.get_object() would be deprecated.

This means the new Reference.peel() without a parameter should behave like Reference.get_object(), and this one should be deprecated.

Yeah, this makes sense.

I am going to be off-line for the rest of the week, so feel free to merge this.

get_object()'s doc now says it's deprecated, so we can remove it in the next large version bump, and meanwhile it just calls peel() with no arguments.