On a high-density display, that 64px x 64px image covers a lot more surface area than 64x64 actual pixels on the physical display. I suspect that this 512x512px requirement is related to scaling factors on devices with high DPI displays.
I'm not even sure why we refer to logical pixels (a.k.a. CSS pixels, etc.) anymore, for exactly the reason you mentioned. We should be saying things like "this only needs to be 512px x 512px because it's never rendered larger than two degrees of viewing angle" or something to that effect. But I guess that's hard to think about... ems could be a good compromise.
> never rendered larger than 64x64
On a high-density display, that 64px x 64px image covers a lot more surface area than 64x64 actual pixels on the physical display. I suspect that this 512x512px requirement is related to scaling factors on devices with high DPI displays.