Concerns (in the broad sense, not ActiveSupport::Concern) can be separated any number of ways. The important part is delineating and formalizing the boundaries between them. For example, a worker running in Puma might instantiate and call three or four or a dozen different service objects all within different engines to accomplish what it needs, but all of that runs in the same Sidekiq thread.
Inserting HTTP or gRPC requests between layers might enforce clean logical boundaries but often what you end up with is a distributed ball of mud that is harder to reason about than a single codebase.
I would also never connect services without a queue unless the message can be discarded (then I can use a pub sub). Using http is one of the most amateurish ways I can imagine to connect two services that I wrote. Even the thought is cringe. Is this common?
You can have a "big, beautiful" Rails monolith codebase used by both Puma and Sidekiq queues etc, and that works well from most standpoints. But RAM usage will be pretty high and limit horizontal scaling.
Fixing this would probably require some changes at the language level, or at the module-loader level for some languages that use one.
Think about it from another angle. RAM is expensive when you're talking about cloud infrastructure at scale.
Or, think about it this way.
For a given compute budget, I can run X instances of Rails, or X times Y instances of a framework that uses less RAM. 100 instances of Rails versus 300 instances of Express or whatever.
Now, it still might be worth it to run Rails, because developer time is more expensive than cloud infrastructure. Or maybe you don't really need to scale. Or you could skip the cloud infrastructure and go bare metal, which is what Basecamp does last I checked.