Memory mapped writes ended up being a very deep rabbit hole and I stumbled into more silicon bugs/quirks. For now I'm using peripheral mode for writes, but in a cleaner fashion (passing the SFR field directly as an argument vs computing offsets manually), while memory mapping reads.
This should work just fine.
Only three more peripherals to convert over (front panel SPI, SERDES low speed IO, SERDES DRP) and then this whole massive refactoring will be done and I can get back to feature development.