Don't be confused by "high poly" and "not optimized", these are two very different things. Good textures and renders are part of a good model, but topology is paramount also.
BUT, define "too many polygons"...
If you want to create a model for a game engine, you'll have to reduce the polycount and bypass this limitation through textures and normal maps. OK.
But what if you need a production-ready model, which will be animated and rendered in a CPU-based renderer ? Polycount doesn't count there, but photorealism does. And if your model is to be added in a live shot with some close-ups, then all details must be modeled because normal maps won't do the job. For such a shot, a 500k polls model is then acceptable if it renders nicely and behaves correctly when added in comp in the live footage.
The good models that sell, in my humble opinion, sell because they provide a response for a need. If they were flawed, they wouldn't sell, just because disappointed buyers would post negative reviews about them.
But a good model that sells is : a good topology, with nice UVs, nice textures, a level of details defined for a need (game-ready or production-ready), etc. If the model doesn't spare hours of work to the guy buying it, then it won't sell, it's that simple.
As for knowing in advance if a model will sell or not, it's always the leap of faith. Will the design interest buyers ? Will a buyer be in need of what you made ? etc. On the other hand, even if not sold, a good model in a library can showcase your talent and help people interested in other models make their mind about your own skills.
That's why one should always be sure to avoid putting crappy models on a library just to get a higher quantity, because it just doesn't give confidence to a potential buyer.
Then, considering free models, the good ones are not that frequent, and a lot of what you can find on the web for free is likely to have been stolen or extracted from video games. If you want to use them in a commercial render, then it's a big no-no, so there will always be a room for good original models, even it the sales are quite erratic, as Ricardas pointed out.
Finally, I don't buy your idea about high-poly models that students wouldn't use because they are too render-intensive. I have a simple 2010 iMac, and I do all my modeling, texturing and renders and animations on it without too much of a fuss, even in scenes sometimes counting millions of polygons. Even softwares like Vue or Poser can handle quite heavy models nowadays.
Students won't buy expensive models not because their computers are too slow, but because they don't have money for that. If you target a large audience, I don't think CGTrader is the right selling place. You should go for the hobbyists market (e.g. Renderosity, Cornucopia3D, Daz3D...), but penetrating it is very complex, especially since you'll have to go through a harsh beta testing before being admitted to sell there, and hobbyists expect to have the quality of 100$ models for 10-15$. Besides, you'll have to learn how to package the products in a specific way, for them to be Poser or Vue or Daz ready...
If you want to target high-end market, and sell less for more, CGTrader or Turbosquid and equivalents are the way to go, but they require complete products, usable easily in any package (interchange is paramount also) and likely to save freelancers or small studios hours and hours of work. Nobody would buy a 100$ model if it didn't save at least 1 day of work. The same way that nobody would buy a 10$ model that can be done in half an hour.