Writing that made me realise just how unintuitive this UI is. There is room to reduce the learning curve here.
Thinking about what could be done that's hopefully not too involved to implement, wondering if the following would reduce the learning curve:
1) Change the animation. The "sometimes it works, sometimes it doesn't" shrink-to-bottom animation comes across as cosmetic. How about switching to an animation that gives a meaningful hint as to what exactly happened when the drop-down was changed?
For example, a plain pull-down animation, where the newly-selected contextual permission slides down coming from immediately below the "Enable these capabilities when top page matches" drop-down's line (and below that line's z-index), sliding over the top of the previously-selected contextual permission.
Also, making the animation work consistently would help it be perceived as meaningful.
2) Change the text "Enable these capabilities when top page matches" to something more like "Capabilities for when top page matches". Makes it sound less like an active setting, and more like a description of what's shown.
Could these changes be helpful enough and minor enough to be worth considering?