What is the rationale for making them part of the chain, instead of just making their hashes part of that chain, like with git LFS?

This way, you would know about all the messages and blobs, but only download those you want, and match their content against the hash. Of course, this comes with less replication, therefore less availability.