Yes, for that side of the problem content-addressing is a good solution. I've also explored content-addressed RDF and OWL in a past project[0], based on IPFS and multiformats. While it gives you good certainty about stability, migrating semi-popular concepts across the ecosystem is a big pain, as you have to regenerate all dependent data (if you go the hardcore route).

I think package managers are in general a pretty good abstraction that can foster structured collaboration across independent (but communicating) parties. It can also enable standardization-via-usage rather than standardization-via-authority, which I think has always been all too common in the semantic data space.

Plow is at its core also very much based on content-addressing, where after the dependency resolution phase, we mostly deal with the hashes of the package contents rather than their name-based identifiers. That also has the nice side-effects that you could easily host a private artifact storage that retrieves the packages for on-premise usage. The public package index is also hosted in a git repository, so all the building blocks would be there that you can record the whole state of the system in a content-addressable manner if you wanted to.

[0]: https://github.com/rlay-project

I’m not sure what you mean by “regenerated”. Sure, URNs will need to be updated to content-addressed ones, probably gradually, but you’d need to do the same to switch to a package manager.