Well, this is an interesting question!

I'm not sure I'll be able to answer it in a satisfying way. I think we now start getting into the tricky area where we need to pay careful attention to the distinction between meaning and language. In particular, I am wary of making claims about what is "real" because I'm not sure the concept is defined precisely enough. I'm not sure "real" is real.

We can instead say the following. When we produce a mathematical model that encapsulates our knowledge of the behavior of water molecules and find that it correctly predicts our microscopic observations to within experimental error, we in some sense "know how water molecules behave". They behave how they behave whether or not we know it. However, if we are right about how they behave, then it must also be that they form waves under certain conditions. And, lo and behold, we observe water waves.

So--and this is more obvious I think to people who have studied information theory and complexity theory than to philosophers--we're making statements about the correlation structure.

We don't need to imbue the molecules themselves, or the universe, with the phenomenology of "logical consequence". We observe correlations and say things like "da/dt = MG/r^2", and can remain agnostic about how the universe "actually" works, or even whether this is a meaningful question or not. We can retreat to the point where we say that it sure looks to us like it acts like that.

Anyway, the point is that certain beliefs we have entail certain other ones, so that there is no extra information by learning the latter, save that we hadn't already made a mistake.

Now it's possible that we could find egregious violations of the expected correlation structure in water molecules, and these would be enough to stop waves, but we also find other egregious violations that make waves appear again: the universe forces waves to appear despite it complicating the interactions between water molecules to the point where we may not even be able to describe them. That's certainly a possibility!

Crowds of people can be that way: the only way, sometimes, to understand the behavior of a crowd is to understand some high-level goal of the crowd, i.e., the behavior of the individuals is driven by the emergent property, and cannot be understood on a person-by-person basis without that context.

But the amazing thing is that the universe appears not to be that way. Reduction seems to work: when we understand the behavior of the components well enough, they seem to be comparatively simple and local, and larger-scale properties are what we would predict given the small-scale properties.

Now, the above description involved us a lot, trying to figure things out. But our figuring-out did not impact the workings of the system. It still works the way it works (if this even is a coherent way to think about it); the patterns are still there, they still have consequences. In a sense, we didn't do anything.

All we did, with rejecting strong emergence, was to say: you know, we didn't need to learn more stuff. We could have predicted the large-scale observations once we learned the small ones.

Eliminative materialism, in my view, simply says: maybe it's all like that. Maybe after learning all the simple rules, in principle there's nothing else to do but work out their consequences. Consciousness seems different from non-consciousness, but waves seem different from non-waves. Maybe if we understood the fundamental operations of the components of the brain and could reason through the consequences well enough, we would find that, huh, consciousness doesn't require extra beliefs--it just is what we ought to expect given what we said we already believed and how we form additional expectations from existing beliefs.

I hope this has clarified the distinction between our talking about reality (i.e. modeling reality) and talking about our models of reality (i.e. modeling our models). When we say something "cannot fail to accrue" in our models, we're talking about our symbology and representationalism. We're not talking about a constraint on the universe but rather about our prediction for a constraint on the universe (but we can ask it whether we were right or not, in this case, by making observations).