I'm pretty much looking for the same thing.
It's sad how most of these project try to add their own layer ("vocabulary") on top of the relational semantics without first trying to surface as much of it as possible.
What I'd like to see is something that can connect to a pg server (or sqlite, or mariadb, but I think settling on one is more feasible) and tries to surface as much as possible of the underlying data representation, and add features that are translatable to how pg works, instead of just using pg (and in grist's case, sqlite) as a dumb backend.
* it doesn't sync with anything but has its own data store, built around sqlite, and you can indeed download "documents", which are just sqlite files.
(I looked at how they encode summary tables, and it's not as views, so the grist<->sqlite translation is not as shallow/isomorphic as one might hope)
* authentication is not taken care of by grist-core itself, but they provide different integration methods (saml, forwarded headers, … ?)
I tried the `grist-labs/grist-core` docker install: it's quite easy to set up and most of the useful features are there (I think it's mostly snapshots missing).
I don't think "free software cosplay" is totally fair in this case.
And compared to the alternatives (nocodb, mathesar, baserow) they seem to provide the best self-hosted/open source solution in terms of polish and features.
*Edit* I've been looking for a while for a similar software stack for the organisation of a coop I'm a member of, and have not found anything better as of now, though open to alternatives!