The LSTM would be trained against an existing high quality Rust code base and it's IR; so I wouldn't expect a bunch of unsafe literal translations.

Whether the C side IR has enough debug metadata or could be annotated automatically would be the make or break.