Refine with refmac and see if the positive density shows up there too. If it does, then either there is something peculiar about your structure or both programs suffer from the same bug that somehow only shows up with your particular dataset. If not, then it is indeed phenix bug (still rather specific to your dataset).
Although this may definitely give more clues, I see this way of debugging this problem more like treating the appendicitis with Tylenol -:) This may help in some way but will unlikely identify the problem and fix its root. Solving a puzzle like this typically involves reducing the amount of "variables" by simplifying the task that still results in problem rather than making it more complex to solve by involving more "variables". Oh, I guess you know what I mean by "variables" so I don't go into it. All the best! Pavel.