What I want to see in a language would be a bunch of syntactic sugar which corresponds to a MIR like Rust's by a well-specified reduction that is likely to be consistent from build to build.

The reason that this could be really cool is that you could hypothetically hash these MIRs and store these hashes in a public database of open-source code. (You would have to include hashes of whatever functions they call/depend on, and you would have to solve foreign-function imports.)

Then you would have this cool language where modules are simply saying "here's a mapping of names to hashes in the public database," with automatic recognition of "hey, these two functions you're importing from these two modules are actually the same function, I don't need to disambiguate them." If your dependency graph isn't a tree and one of the branches can't cope with a dependency-upgrade, that's fine too -- you upgrade the dependency on the one branch only. The point is, all of your dependency-semantics are now reduced to a really straightforward reasoning about unique names.