I'll defer to @carlosmn and @ethomson.

This PR seems fine to me. The refs are loaded in local_connect, so it makes sense that they should be unloaded in local_close.

I think eventually we should go further with fixing this because right now it appears to me that calling local_connect doesn't make any effort to check if the transport is already connected and will happily reallocate all the structures and simply overwrite the existing allocated objects. It would not be hard to make it an error or a no-op if you try to connect a transport that is already connected.

However, that goal need not be part of this PR. This fixes the memory leak from closing and then reconnecting the transport.

Thanks @arrbee!

The remote reference list needs to survive a disconnect. Otherwise you won't be able to call git_remote_ls after you've disconnected from the remote.

Hmm, okay. Then I think we need to put some checks in place (as I mentioned above) to use the internal state of the transport to avoid leaking memory. We can write an alternate fix for this tomorrow.