One thing that IBM did that really toasted me back
then was messing
up on the 8237 DMA controller hookup such that memory-to-memory DMA
didn't work. It could have made the whole business of extended
memory use a lot simpler.
I still think they should have used the 8089 'I/O coprocessor' which was
essentially a fancy DMA chip. That would have, for example, allowed DMA
transders to cross page bounadaries.
OK, there necver was an 80289 (or whatever), hut you can bet Intel would
have made one if there'd been a demand (read : It would have gone into
the IBM PC/AT).