Very cool! Was this at all connected to the recent discussion about converting SMB to use a mapper? When I get my new MMC3 dev board in I'm going to have to give this a go 
Very nice. 
When I did it, I added PRG-RAM and uploaded some subroutines into it, to move them out of the main bank. Not necessarily the best way to do it, but it worked. 
With some extra PRG space now, I'm wondering if it'd be worthwhile to move the titlescreen data out of CHR-ROM and rewrite the routine to load it from PRG. It'd free up some tiles.
When I did it, I added PRG-RAM and uploaded some subroutines into it, to move them out of the main bank. Not necessarily the best way to do it, but it worked. With some extra PRG space now, I'm wondering if it'd be worthwhile to move the titlescreen data out of CHR-ROM and rewrite the routine to load it from PRG. It'd free up some tiles.
Well with MMC3 you could always use bankswapping to get more CHR space.
But then you'd have to rearrange the existing tiles in CHR ROM to make room, which means deciding what tiles can and can't appear on the screen together. I can see splitting the blocks associated with normal, castle, and underwater environments, but what else can be split?
You could just prep 4K chunks for GB and SP patterns that contain your alternate tiles and bank them in based on level. That way you replace only what you need to.
ROM hacks don't have to be very practical after all
ROM hacks don't have to be very practical after all