The original idea of the VIRTUAL class, from which VIRTUAL_ENTRY etc are derived, is that it acts like a real ENTRY, so for example a citation of an earlier lab result as justification for a diagnosis is in the data as a VIRTUAL_ENTRY corresponding to the cited OBSERVATION . Rendering software knows it is a virtual entry so can show its contents but in a way that makes it clear it is a citation. Similarly, the querying service can ignore virtual Entries, Compositions etc easily, because they are not instances of the primary classes. But a special query could find citations as well.
There are some diagrams of use cases in earlier discussion on this topic. BMI example here and Care plan example here.
At the archetype modelling level, you would model original content normally, and you would constrain virtual items within citing Compositions (such as referrals, discharge summaries, reports etc) to be of certain types if you wanted.
I don’t claim all this is 100% worked out, at least I have not had the time to work on it for a few years, but I think the main ideas are in place. I hope to get back to this and related questions soon.