What the OP writes about is a good first step and frankly, is very close to right, but why stop there? Why not ditch ActiveRecord models all together and totally decouple persistance from the "app"? Honestly, the coupling of your models and logic to a database is really where most of this trouble starts.
In the Obvious architecture the app is separate from the delivery mechanism (rails/sinatra/whatever) and also from the external persistance mechanisms like databases, queues, caching, and so on. You can mix and match both delivery mechanism and persistance methods. So, you could write a CLI app against the filesystem and then make a rails web app with a mysql persistance mechanism without changing your "app" at all.
The whole thing is built from day one to be TDD and in a recent app I wrote, it has some 78 rspec tests that run in 0.03 seconds.
I'm in the process of documenting and open sourcing the whole thing. There is a project generator gem available now called 'obvious' and you can read more at http://obvious.retromocha.com/
When I have better documentation and examples available, I plan on posting to HN, but all questions are welcome.