My hypothesis is that, at least on VisionOS, some apps are full of — almost cluttered with — 3D objects; and so Apple felt that, for the sake of your eye being easily able to jump to "where the UI is" amongst all that, the user needed to be able to visually differentiate/distinguish action buttons (incl. "buttons that launch apps" — essentially what these app icons are, esp. on the mobile OSes) from those 3D objects. This was achieved by ensuring that action buttons are always button-shaped, rather than allowing them to be arbitrary-object-shaped.

Note that, in this UX-design paradigm, the icon on (in?) a button still can be its own standalone object of arbitrary shape, rather than being forced to be button-shaped itself (see e.g. the Stickies or Game Center icons in TFA.) But that standalone object has to then be "encased" in the "app button" glass (as if encasing something in a puck of pourable resin), to make it visually obvious that this object is functionally a button, rather than just being some random 3D object with its own arbitrary interaction semantics.

Funny enough, this is almost exactly the complement to the problem of visually differentiating action buttons from 2D content. In a 2D UI, you want to make the action buttons more 3D-looking than the 2D stuff around them, to help them stand out. Thus the Windows XP / macOS 9 era of "jelly" buttons with that visually bulge toward the screen — standing proud of the content, affording touch.

But if everything is 3D / stands proud in arbitrary ways, then overlaid actions will stand out better if they're less 3D — making it clear that they're sitting "on the HUD" rather than "in the world." Such objects can be literal 2D — or you can get fancy and choose some unusual middle-ground, like the sort of 2.5D papercut-diorama look that "liquid glass" achieves.