I still think that generating a Zig-Rust transpiler would be a better approach, given all the LLM quirks, including the ability to just /goal the model with binary-identical LLVM bytecode.
However, an open-sourced tool like that would've greatly harmed the Zig ecosystem and community.