I share this as my experience only, not what you should necessarily do. That’s only for you to determine.
As a result of using Starling, I made a habit of only loading resources at runtime. If resources were embedded, effectively twice the resources were consumed: Once to load the main file (containing embedded assets), and again once those assets were uploaded to the GPU. Loading assets at runtime simply uses less resources.
When using ATF assets, such as spritesheets, these are encoded for the GPU. The compression it supports is handled by the GPU and so they exist in VRAM in a compressed state, consuming less VRAM! (but note, you can’t use the optional JXR compression with OpenFL due to license restrictions).
I hope this starts to answer your next question:
I provided a bit more info on achieving that here.