It's non-trivial to implement in a general way. It would require fairly deep rework of all the final-form transforms to produce a mapping from source elements that have associated keys to their deliverable-specific anchors. It also requires some mechanism for capturing the configuration details for the associations between a given root map and the deliverables generated from it, including the filtering conditions, any processor-specific parameters, and the time when the deliverable was generated. That is, you need some kind of "publication manager" functionality that at least provides the framework by which you can coordinate the publication details among a set of related publications. Basically, for all the publications in a set of publications that link to each other, you need to publish each one at least twice so that each deliverable for a given publication accurately reflects the anchors in the target deliverables from the other publications.
In the context of a local environment where your implementation doesn't need to be so general the implementation can be easier because you can impose constraints or rely on conventions that simplify the problem. For example, if you don't use key scopes or branch filtering the problem becomes easier (because you assume a simple deterministic mapping from source keys and IDs or filenames to deliverable anchors, which you can't do in the face of key scopes or branch filtering).
Conceptually you have to maintain, for each root map from which a deliverable is generated, a set of mappings from source elements with associated keys and the deliverable-specific anchors they become so that you can then map references to those keys from other publications to the corresponding anchors (which also requires that you have a way to deciding which deliverable to link to for a given deliverable being linked from).
These mappings could be maintained as sets of key definitions that replace the original source key bindings with bindings to the actual deliverables (using @scope = external and whatever @format value is appropriate) or they could be in a more traditional database of some sort. But you can completely define the cross-deliverable processing in terms of replacing sets of key definitions that point to the target publication's source with key definitions that point to anchors in a deliverable generated from the target publication.
I've been waiting for two things to happen before trying to implement this myself:
1. The OT 3.x code to stablize, in particular the preprocessing pipeline.
2. For a client to fund some or all of the work
1 is pretty close, I think. 2 I'm still waiting for, although I have had a couple of inquiries in the last month or so that may or may not pan out.
Cheers,
E.