First as the GUI discussion (and at least C/C++ interop) is off-topic I point to:
I’m not sure it’s wanted to add to that thread (nor start a new one?). Since the GUI thread within this thread is more general than only a “Chess game” and people may be interested (at least if there’s new info here, e.g. relevant to Julia 1.0) I propose splitting it out (I just don’t know how to extract the discussion, and I believe I don’t have the permission).
A more closer to home example of wrapper on a wrapper may be, PyPlot.jl (that wraps matplotlib), and is at least graphical.
There’s no big problem with wrapper on wrapper, across languages (at least to C/C++), or within, e.g. Plots.jl to PyPlot.jl [to matplotlib].
While PyCall.jl README doc mentions PyQt you could use (may not be wanted because of GPL, and only for Qt4), it also mentiones the better option (at least to get LGPL) would be:
“PySide2 (part of Qt for Python) 5.11.2 Technical preview / 21 September”
Yes, Knet and/or Flux may be better for machine learning, but not because of wrapper issues. TensorFlow.jl even claims better API than the original Python one. But now we’re even more off-topic.
I don’t think you want to work with the framework of e.g. Windows, as then you’re no longer portable (unless those, e.g. from C#, not the “C”/lowest-level APIs, are better).
You could call the lowest level of the OS, for GUI, or non-GUI work. That’s bad for many reasons, and would tie you to Windows only. For GUIs (or non-GUI stuff), even if you accept being tied to Windows there are better APIs/frameworks than this (and despite the other side of the API written in C, it’s really just compiled machine code, and kind of wrong to say the interface is C, despite the ccall keyword used to call it):