Thanks for providing an example, it helps a lot !
As you guessed, the problem is in the renderer in the part_feedback function.
For the adaptive behaviour no problem because we are in fact not using adaptive but using Tim's adaptive for multi part questions behaviour, so lines 434-442 $showfeedback is corrected according to the part graded state, not the question graded state.
Unfortunately we can't do that for other behaviours as all we have is the question graded state.
A very quick fix would be to show the feedback if the question graded state is wrong or partially correct.
Of course this would not be perfect as feedback for some parts will be displayed even is that part's response is correct, but at least if all the question is correct no part feedback will be displayed (only general and combined feedback).
I think this is the best we can do without too much work (we don't have enough information provided by the behaviour to know the graded state of the part). But maybe I am wrong ?