Did you meant to use tab-width: 20?

Shouldn't we set mFailed = true; then return false; here?

Is this just for debugging? If so should it be ifdef'd somehow?

no else after return probably.

no else after return.

Okay, that explains a lot of the choices here :)

Can you make sure to document that, so that people sending patches here know that it's a bit different from the rest of the codebase?

I don't think you intend to expose this to consumers more generally, so probably just forward-declare them here and move the code to the cpp file.

nit: redundant, it's already private.

If you're using the STL, why not just returning an std::unique_ptr<WaylandProxy> here?

Also it might be worth having some documentation on how is this expected to be used.

Similarly, this could probably be a static std::unique_ptr<WaylandProxy> Create(); in the proxy class.

It seems these two functions could just be methods in WaylandProxy. In fact, what is the difference between WaylandProxyRunChildApplication(proxy, args) and proxy->RunChildApplication(args)?

Maybe WaylandProxySetVerbose or something would be a bit clearer.

const probably.

const

Maybe we should have some sort of arena / freelist for these for performance, if heap-allocation here is hurting us.

Maybe worth noting here that resize() is guaranteed to not invalidate mData, as per:

In case the container shrinks, all iterators, pointers and references to elements that have not been removed remain valid after the resize and refer to the same elements they were referring to before the call.

In https://cplusplus.com/reference/vector/vector/resize/

Should these be >=? I guess it's unlikely to get a 0 fd, but maybe possible if someone closes stdin / stdout / stderr?

!mToApplicationQueue.empty()

!mToCompositorQueue.empty()

This is more expensive than it needs to be because it shifts all the elements in the vector. Let's use an std::deque instead of an std::vector?

You can use auto here if you want.

if (mConnections.empty())

Probably be explicit about the memory orderings here, at the very least, this should be acquire, and the = false should use release.

More documentation on these would be nice, either here or in the cpp file.