Mainmatter head of engineering here. From what we've seen on real projects, I can confirm that asking an AI to migrate the code results at best in somewhat functional but unmaintainable code, and starting from scratch is a better approach… However, starting from scratch doesn't work if the legacy codebase is actively being worked on (eg bux fixes and new features and refactorings) or lacks test coverage and documentation.

Starting from scratch also implies one major release where suddenly it's all Rust (so very risky). In contrast, we found that it's a much smoother transition for end-users and maintainers when we approach large codebases progressively, module by module, while the C engineers continue working on the C code and learn Rust in parallel.

Of course, each migration is different but I wanted to clarify with our field experience.