Every time I've used multiple decoupled libraries instead of one large framework, things have gone fine. If a particular library isn't working out, I just swap it out for an alternative (or roll my own, if worst comes to worst) and continue building.
Every time I've used a large framework, I've ended up with major buyer's remorse. There's always at least one particular piece (like routing in your case) that's a bad fit, but everything is so tightly coupled, I can't just swap out the offending component without causing problems elsewhere...so I often end up accepting a bad fit.
For me, the productivity gains of big frameworks have always been outweighed by drawbacks as soon as I hit nonstandard use cases. I always find myself wishing I'd used multiple smaller, decoupled libraries instead.
As far as Ember Data goes, this is something that it's clear that a lot of people do indeed want. That said, you can use Ember perfectly well without Ember Data (see Discourse for example).
It's funny that you point to Rails as an example of consensus, because when you look at the server-side development community as a whole, there is certainly not consensus that the Rails way is correct. I wouldn't expect complete consensus on the client side any more than I would expect complete consensus on the server side. Rails (and others) have shown that you can have multiple healthy ecosystems each with their own points of consensus.
You're right about ember-data being optional but I mentioned it because it's what we were using on the project I was on.
I guess what I meant when mentioning rails was that there's a lot less change happening on the server-side than on the client-side these days so it's easier to confidently build an opinionated framework there. It's hard to favor convention over configuration when the very platforms you're targeting don't have established, agreed upon conventions.
Data persistence on the client is another area I feel hasn't been nailed down in general and thus can't support a lot of opinionated assumptions. Should data be persisted at all? Ember caches data from all requests to the server in a "store", but it's not clear how you keep that cache up-to-date. Perhaps you could use websockets to keep resources current but at the time I was using Ember there wasn't a clear, well supported way to do that. I think the jury is still out on whether we should be duplicating our schema on the front-end at all but Ember-Data has gone all-in on mimicking ActiveRecord. Again, the dust has not settled in a lot of areas of the front-end development world as far as I'm concerned so I don't feel comfortable with these decisions being made for me.
I tend to lean toward convention over configuration. On the server I'm a big Rails fan but I feel like it's solving problems that aren't moving targets. On the front end I prefer Angular because it solves the main headaches (two-way binding, testability, modularity) of front-end development without being too opinionated about the stuff we haven't reached consensus on.