I like your point about users "reacting" to an limited range of choices, explicitly stated (and surely using domain specific terminology). It's a grammar you don't have to internalize, with libraries you don't have to know about, or know the search terms to find term. It's very "discoverable", and the limited and clear range of choices minimizes the information content (increasing the probability of obtaining the desired outcome). I'm sure it would also prune the tree by omitting choice combinations that don't make sense.
I think people don't like the system guessing for you, like Office's Clippy. Maybe it's OK if the user asks for the guess - like getting a list of search results; or vim/bash completion; or suggestions (google, Stackoverflow)
What's important, though, is that these people don't build UIs. First and foremost they manipulate or analyse data. The data is at the center. The important thing is a simple data model that follows the closure principle (filter tables and you get tables, filter text and you get text). The UI is there to support this simple data model and it has to be similarly reductionist I think. Placing arbitrary UI widgets on forms is not the way to do this.
Yes, more fundamental research please :-) Maybe machine learning could be used to let users correct the system's guesses rather than initiating some action. That would go further in the direction that graphical UIs have begun, which is to let users select from what is offered to them and react instead of making them give commands to the system.
Or maybe some of the thinking that has gone into computer games would be useful, I don't know.