Yeah, I talked to grinvader & kode54 about it while I was waiting for the post to be approved, and they made many of the same points you did. I'm not really suited for discussing the PPU stuff, though, so I'll take your (and their) word for it.
Beyond that...I think the first version would be limited to just PNG decoding, with an opaque scratch buffer for the decoder itself and another 1-2k of RAM for the game to put PNG data into for processing—I'd have to experiment with 16-bit PNGs to see if this is overkill or not (2k almost certainly will be, but not sure about 1k), though.
(ETA: Okay. Preliminary testing with 8-bit PNGs indicates 1024 bytes for shunting data to the chip isn't enough—48x48 exceeds it. I might actually suggest going as high as 4k bytes of RAM.)
(ETA 2: ...I should've stuck with my original thought of 16k RAM, though double that would probably be safer. 256x256 (actually 256x142, but meh) ends up being over 9k, optimized.)