Actually one of the things I found during my work on transducers is that actually they seem to me more implementable in C++ than in Haskell. For example, you can implement state-less transducers with almost zero overhead, something that in Clojure is definitely not true (so the standard transducers do hide state inside the closure of the reducing functions). It is also easier to do n-ary transducers for zipping and unzipping and everything, as you intuit, is type checked.
Still, I agree that transducers are most useful when you have libraries that make sense of them. For example, you could implement a lot of "Rx" using transducers. Ideally I see a user never writing a transducer, and occasionally passing one to some library that provides "higher order" reactive collections (channels, observables, etc.) that can be transformed with transducers. For example, my motivation to implement them in C++ was to build "observable/reactive lenses" (something like "cursors" in Om) [1] where you could use transducers to specify how you "zoom" into the virtual sequence of state values...
Anyways, they are an interesting coding tool, but they don't really influence the architecture of the application.
[1] atria::funken https://github.com/Ableton/atria