I agree with most of what you say. I'm a long-time fan of calculating more complex things client side, as you can see from my climate model (currently all calcs within web browser, evolved from java applet to scalajs).
Also, in regarding social media, keeping the data client side could make the network more resilient in autocratic countries (many), and thelp this become truly a global alternative.
On the other hand, some 'trunk' server interactions could also doing more not less, bundling many 'activity' messages together for efficiency - especially to reduce the duplication of meta-info headers in clunky json, and work of authentification-checking (which I suppose has to happen to propagate every upvote in Lemmy?).

[–] rglullis@communick.news 2 points 5 hours ago

bundling many ‘activity’ messages together for efficiency - especially to reduce the duplication of meta-info headers in clunky json

Seems like an optimization that is not really needed. The data format is not really the bottleneck, there are ActivityPub relays that can send messages in bulk and ActivityPub is built on LinkedData, which means that there plenty of powerful libraries in most languages that can parse and produce JSON in a way that keeps application developers with a consistent semantics. The more people try to change the data format in the sake of "efficiency", the less portable and useful it would be.

and work of authentification-checking (which I suppose has to happen to propagate every upvote in Lemmy?)

Yes and no. Most of the current software do authentication by using HTTP Message Signatures, so after you fetch the actor's public key every request is authenticated by seeing an HTTP header, which makes it no different most common authentication schemes.