One of the interesting-to-me realizations is that these marker protocols are vaguely attribute-like (since they are compile time only). The problem with this viewpoint is that we want ConcurrentValue to participate in the generics system, and making attributes do that would open a ton of complexity we don't want.

This proposal introduces the concept of a “marker” protocol, which indicates that the protocol has some semantic property but is entirely a compile-time notion that does not have any impact at runtime.