Thanks for the quick reply!
Hm - am I right that the Trello card is proposing something slightly different?
Seems like the proposal is to keep the texture as PNG compressed bytes on disk, do the decoding in memory - then upload the raw uncompressed bytes to the GPU?
That would appear to result in a smaller on-disk size (and less time fetching from disk) at the trade off of some CPU time to do the decompression. It also means you've got the memory overhead of the truecolor texture in memory.
I think what I'm suggesting is a tiny bit different in that I'd like to continue leaving the RGB channels as PVR compressed (as I'm seeing acceptable quality there) - which gives a reasonable size and memory pressure - but upload a second texture containing true-color alpha.
Forgive my ignorance here - I could be totally wrong