Yeah, to be clear: pure implementations of CSP are probably useless but "near" implementations are abound. I don't know Go, so Erlang came to mind... But it's definitely a little way from pure CSP.
I suspect most "pure" models are nigh useless on their own. They all require a bit of compromise to interact with the world, or exist as a subset of some other language/tool. However, they can be awesome for describing in specifications or analyses of systems. And when languages support them in some fashion (even if impure) it allows us to reduce the semantic distance between specification/requiremnets and implementation to a point where we can be confident that we've done things correctly.
Haskell is an interesting playground for this since it tries so hard to make "pure" practical. Interestingly, they achieve this to a large degree using monads. I think that idea is fairly generalizable, though, that pure formalisms can embed impure ones as models and then have those be executed by an RTS.