Well from my perspective question number 1 has a flawed premise, question number 2 would be a pretty boring answer and question 3 is kind of confusing and abstract.

But since you didn't get anything from anyone else yet, I'll have a go anyway.

1) There isn't one element of a piece of middleware that needs to be done right. All elements of all pieces of middleware need to be done right. If any element of any piece of middleware isn't done right, I will use something else or write my own. However, compromising a little on this, I'm going to say that stability, complete lack of bugs (and prompt fixing of the ones that come up because complete lack of bugs does not exist) and generally doing what it is supposed to do flawlessly is more important to me than the interface or the speed.

2) Best experiences? Blitz3D - spectacularly stable and solid 3D engine in it's time. Bass sound engine - very stable, very clean interface, does what it says on the tin. Newton physics library - others may be faster and have more features but Newton is really, really tight. I think the right word is deterministic? Essentially that with the exact same inputs Newton will produce the exact same results. Very useful if you need that kind of accuracy.

3) I don't really see how A and B relate. They don't seem to be mutually exclusive or complementary in particular. Anyways, A) sounds very good. If I didn't want maximum output from minimum input, I wouldn't be using middleware. That's the point of middleware, to save me as much time as possible. B) I don't really get in an abstract context like that. Perhaps if I had a concrete example of how a 3D engine would be modular or how a sound or physics engine would be modular, but it doesn't really seem like a big deal to me from the abstract description.

I suspect that's not terribly helpful because I've kind of taken issue with your questions, but hopefully it's better than nothing.