The criticism raised here only applies to REQ/REP socket types. They're introduced early in the literature (i.e. the zguide [1]) because they're easy to reason about. But in my experience I haven't seen them used in systems because they're so limiting. Not just due to disconnects, but because a REP socket can only handle a single request at a time, so therefore concurrent overlapping requests throw state machine errors. Thankfully the logic in the REQ/REP sockets is very simple, so by not using them, you're not losing much.
[1] http://zguide.zeromq.org/page:all