It's only ugly because it was meant to be a federation of different programming paradigms. You can mix low level C calls with homegrown RAII frameworks, or mix traditional OOP with functional programming. Once you start throwing in preprocessor macros, meta programming, and templates, you can have a codebase that is incredibly complex to understand and maintain.

As for the header/implementation separation, I always thought it was a good idea to have that flexibility. I've worked in codebases where the compilation units were fairly large and complex and would be implemented in separate files. This would be similar to partials in C# I've also worked with some code that would determine the target compilation units at build time. Not saying it was the best approach at the time, but it was one option.

If anything scares off people from C++, it would definitely be the breadth of the language, though. You could work in C++ for over ten years and possibly not even encounter or use half of the functionality it provides. And because of it's breadth, going from one C++ codebase to another could look completely different. Especially when looking at something written for a Unix/Linux/Posix system vs Windows.