The XZ incident would have been so much worse if it would have involved a collision, pushing one object variant to github.com, and one variant to git.tukaani.org.

Then you would have security researchers making conflicting claims depending on which repository they first pulled from, even though they are on the same git commit hash.

In GitHub, fork networks are represented as one repo on disk (this is the main optimization that makes community pull requests possible at all). So you could fork a repo and then push an object with a hash collision to it to change a file in the original, in theory.

(In practice this is harder, as the article mentions, because the new forged object would have to be a valid gzipped git object of the same length. And GitHub probably knows about this type of attack and might just, for example, prevent existing objects from being overwritten)

Hack into the system holding a trusted repository, and swap in your variant with same signature?

Social engineering?

DNS cache poisoning?