The psql process even exits with an error code 2, which might be not that expected. We could stop reading the file and reset connection afterwards, but this is probably not that easy to achieve (think of nested \i calls).
Well, if you stop reading from the file then I don't think more \i's will matter, no? I'd certainly like to see improvement here, because the difference in behavior with \i is annoying.
What I mean is you need a longjump out of the innermost \i back to the toplevel interactive prompt. This might be not a problem since this is what already happens upon receiving SIGINT, I believe.
On the bigger question of how to better protect all these cases (cut & paste, etc), I think the only robust way to do that is for psql to track intended transaction status itself. With the amount of parsing it's already doing, maybe that wouldn't be that hard to add. It looks like this might require extra libpq calls to constantly check in on server status; I'm a bit surprised that result objects don't include that info.
This doesn't have to be solely about transaction status, though for something immediately destructive such as DELETE or UPDATE one should expect a transaction guard. But for example, pasting something like the following two lines
SET search_path = 'fancy_new_schema', 'old_boring_schema', public;
SELECT * FROM get_item_ids_to_delete(...);
can give slightly different results depending on whether the first statement had it effect or not. And since psql is trying to be very helpful here by resetting the connection, it makes it all too easy to overlook the problem.
What do you think about trying to read everything we can from the terminal using non-blocking IO and only if that gives EWOULDBLOCK, starting the interactive prompt?