The buttons on that type of radio (and on cassette walkmans) have other, tactile indicators of behavior: when pressing one button (before you have even pressed hard enough for it to engage), the others react in a way that hints at the mechanism that controls all of them — the entire set shifts together. In other words, you do not need to use the ‘radio buttons’ on a boom-box or cassette walkman to know that they have this relationship: you merely need to touch them.
These hints are not available on radio button widgets, which have no tactility. Instead, we must learn that some buttons behave in this way. Since every UI toolkit has a distinct way of differentiating radio buttons from checkboxes, and since no non-developer knows the term ‘radio button’, the amount of experience required to generalize from the behavior of a particular collection of widgets to a whole class of different-looking widgets in different applications is substantial. Not only that, but we must actually interact with — and activate — these widgets to even notice it.
Predictability is a much lower bar for a single application than for a UI style in general. This is why I don’t consider a unified style particularly effective, when it comes to fostering it.