Working from shock__'s suggestion, I made a real Frankenstein out of one ROM module for testing purposes. The victim now has a number of wires soldered to it, that run to a mini-breadboard. The breadboard contains a de-multiplexer, which serves as "glue", making a dual-2-MiB module look like a single-bank 4 MiB ROM to the GUS.
With a little massaging of the ROM data, the split looks seamless to the InterWave, and preliminary testing shows some promise. Given that the glue adds to the propagation delay, I have some concerns about corruption. I will try dumping the contents and comparing to the source image repeatedly, when time permits. My first try produced a perfect result on comparison, but that doesn't eliminate the possibility of occasional corruption.
As you may have guessed, the purpose of this experiment is to further test Mega-Em and UltraMID. To that end, I have used only the original GUS patch set, converted to 8-bit. We'll see if it is better-behaved when no additional percussion banks are present, but I have my doubts.
More information to follow.