I'm also building a local first editor and rolling my own CRDTs. There are enormous challenges to make it work. For example the storage size issue mentioned in the blog, I end up using with yjs' approach which only increase the clock for upsertion, and for deletion remove the content and only remain deleted item ids which can be efficiently compressed since most ids are continuous.
In case you missed it and it's relevant, there was an automerge v3 announcement posted the other day here which claimed some nice compression numbers as well
As far as I know, automerge is using DAG history log and garbage collecting by comparing the version clock heads of 2 clients. That is different than yjs. I have not followed their compression approach in v3 yet, will check if having time.