My opinion on this is that there isn’t a satisfiable solution. You can’t have both 1) low impact on existing clients, and 2) any guarantees about maintenance of this information. These are strictly contradictory.
LLVM IR tried a middle ground and doesn’t satisfy either of these approaches. Existing MLIR clients will (rightfully IMO) push back against impact on them because (by definition) no existing clients actually use this information.
My personal recommendation is to pursue the “least-invasive approach even if flawed” approach, and if there develop important clients that can justify why the new thing is required and not good enough, and if there are no satisfactory alternative implementation approaches, then more invasive approaches could be considered.
This is a bit of “boiling the frog” approach to infrastructure evolution.
-Chris