I've also found it's extremely hard to predict ahead of time what projects will turn out to be long lived and large scale so I try to build the simplest thing possible first and then take advantage of the flexibility of modern stacks to iterate and refactor from there. Things like named scopes in rails should be more than enough to keep your controllers from getting too tightly intertwined with your models.
I think predicting the length and scale of projects likely depends on who you work for and what products they want. I might be biased because I've spent most of my professional life working for a large-ish business where products in general end up being reused a lot by other internal products and all have an active life of at least 3 to 5 years, during which different developers will 'own' them. In this case, you can understand how planning ahead might be necessary. I'm currently supporting an application that's 14 years old and small leaks like that end up hurting a whole lot.
Then I guess I'm not as strict for home projects where I'll be the sole owner and developer, but that's the opposite end of the spectrum.
The interesting and frustrating thing about building software is that the answer to most questions is "it depends". What you're describing may be appropriate in your case. In the case of most of the projects I've been involved with adding any significant extra abstraction on top of reasonable use of an ORM would be overkill in most places.
If you need a special data store for testing, just point the ORM at a test database. You don't need a half-assed database simulator when you can have as many real databases as you need.
Then I think Groxx hit the nail on the head with his comment (http://news.ycombinator.com/item?id=1636322). The advice is mostly valid for large-scale projects, things you expect to stay forever, which is what I'm maintaining at the moment; Not for experimental or short-lived code (or whatever is logical on that spectrum).