The moment you ask for reliability it stops being simple. Stream Management + Stanza Ids isn't simple, and it shows - XMPP is the only thing I use today where it occasionally happens that a message isn't delivered and I'm not even notified about it.
On a side note I don't believe that this can't be fixed. Maybe if we had end clients checking directly with each other what's been delivered and what hasn't, like actual humans end up doing anyway, I believe it'd be more reliable. Besides instead of relying on the c2s-s2s-s2c cooperation chain to work, it'd rely on just c2c cooperation, giving much fewer variations to test. Just to clarify I'm not talking about a p2p connection but rather e.g. checking the number of stanzas not with the server, but with the interlocutor's client, through the normal XMPP means, like <iq/>s maybe