Yes, OS in upper RAM, interrupt vectors free (for userspace programs to use!), one arbitrary "main program" loaded into low memory running "in foreground", not relocable and with arbitrary instructions, just like on your average '51, plus any number of resident programs written in a special way, loaded just below the OS and using interrupts to "run in the background".

You want a bondage and discipline system, which could grow and breed hundreds of revisions and still run the same software, just slower and slower with each revision, due to patches, options and version branches added to the API. I doubt more than 2-3 revisions will be ever released (and more than 100 pieces built). The abstraction layer you plan on adding will definitely slow system down. I'd rather think of outsourcing the compatibility issues to external chips, building a mostly fixed memory map and keeping the OS out of the hardware versioning business. Don't plan a PC. Buy an Atari on eBay, play a little with its OS and software and see what I mean.