Anyway, the third undocumented bug is that --single-transactions gets to send its COMMIT even if ON_ERROR_STOP takes hold before the end of the script. I imagined it such that only if every statement in the "-f <script>" was called would the COMMIT be issued - thus the error_stop would supercede and leave the session uncommitted and by default rolledback.
Not seeing the bug. --single-transaction wraps the entire script in BEGIN/COMMIT, ON_ERROR_STOP stops 'processing' the command, nothing in there about stopping transaction or rollback. So the failed \i stops the script from processing anything after that and the session goes directly to the COMMIT. If you want to deal with transactions there is ON_ERROR_ROLLBACK. Though I did find something interesting about that, which will subject of another post.
Then --single-transaction has nothing to do with the script file at-all. It should be documented as issuing a BEGIN at session connect and a COMMIT just before session disconnect - regardless of whether the named script executes to completion, which can happen if it is combined with ON_ERROR_STOP.
FWIW
The way this is written currently I am imagining something like this happens: