Description
Noting down some thoughts after @autoencoder's lead:
The current design uses a relatively conventional notion of "identity" for both artefacts and entities, which means a real canonical, platonic form, incorruptible representation of each Work exists, and when merges happen (and metadata in general is added), they happen against this one identity. One downside is that unmerges can only go back to the point of the merge; there is also no obvious way to model different notions of "sameness" with this design.
Another approach may be to adopt a situational, contextual identity where data is added, whenever possible, to a concrete representation of the Work, potentially without bound in granularity: maybe as specific as a particular playback of a song, a print in a specific gallery, etc. These aspects of a Work's identity can then be merged at read time, depending on the query's needs.
This can potentially be implemented with an overlay layer on top of the chains, but the multi-chain abstraction itself becomes less suitable for a system consisting of many small objects. Needs more thought.