I recently discovered this repo which compiles arbitrary code into a 10 assembly instruction program that loops.
It achieves this by offloading the majority of the code logic to a blob of read-write non-executable data.
https://github.com/xoreaxeaxeax/reductio
You could prove the inputs for each iteration of the loop outputs the inputs for the next iteration of the loop. This is highly parallelisable and the polynomials involved would be tiny making inversion steps much simpler.
You would then need some way to succinctly aggregate all those mini proofs.
Is this pure silliness or might there be something here?