Yes! There is some work being done in this space already; by the incredible people at North-Eastern. One of the more interesting parts of this problem is what you do at the /boundaries/ of the different languages with differing semantics such that you don't accidentally introduce bad lemmas.
There are a few published papers which study the boundary between dynamic and static type-systems, but the hope was to extend it into gradual-typing between different type-systems .
I think I know some of these papers, but could you please point me to some of the references you have mentioned?
Of course — this paper talks about the mechanism which Racket explores to keep the boundary safe [1]. This
paper [2] elaborates on some of the performance challenges. 'Shallow typing' by Ben Greenman et al.
is available only by video [3] and slideshow form [4]. Finally, there's a really interesting paper which came out
this year which uses a verifier to remove the some of the runtime cost [5] of sound gradual typing.
[1] https://www.cs.utah.edu/plt/publications/oopsla12-stff.pdf
[2] https://www2.ccs.neu.edu/racket/pubs/popl16-tfgnvf.pdf
[3] https://www.youtube.com/watch?v=A4AA8nweYAg
[4] https://www.ccs.neu.edu/~types/resources/talks/g-rcon-2020.p...
So, instead of programming in Turing-complete languages, one creates many specialized DSLs with just enough expressiveness. And, in parallel, some automated formal methods. The combination of a DSL and SAT/SMT makes this feasible, which contrasts with automated theorem proving on Turing-complete languages.
Racket seems like a good platform to do this.