Agree with all.
But if 80% of your code can be performant with "ideomatic' code, 15% needing a bit more attention (such as not making use of certain abstractions, like link-to-objects, etc.), with the other 5% requiring diligent attention (Span, etc.) - then your making gradual perf gains as you wish, all within the same language, runtime, and memory pool. This is much better than a re-write and dealing with FFI, or just as likely, full-on serialization between your old code and Rust code.
With that said, there isn't much of an excuse for non-0-cost enumerators, etc. -- given that refactoring tools switch between foreach, linq, indexing or ptr/spanning, you'd think the compiler-chain or jitter could do the same.
Roslyn is a better back-end for meta-programming, but it's too bad MS denies us the benefits of surfacing meta-programming to the language or at least tool level, or we'd have tools that could easily convert linq to highly optimized code.
You could also have compiler-enforced "levels" of C# code that had restricted features/semantics, such as "alloc-free", etc. - right now the choice is just "default" or "unsafe".