There's a documented memory model (https://github.com/dotnet/runtime/blob/main/docs/design/spec...), does that not address this concern?

So, the answer is that I've read that and I wasn't as reassured as I'd like. ECMA definitely isn't enough as it acknowledges. The platforms which exist today are fine. And future platforms? Well we're told it will be difficult to change these assumptions. Yeah, it would be difficult.

For a comparison, the x86 has what that document calls TSO, a very strict "free" ordering (in fact you pay all the time, but, you can't opt out so in that sense it's free to get this ordering on Intel) so 1990s C++ written for Windows just assumes volatile means you get memory ordering -- even though that's not what that means. If you compile brand new code for x86 on Microsoft's compilers today you get the exact same promise, but if you target their ARM platforms you don't get that because it would be expensive so, too bad.

It depends on what it says?