I think in practice PGO for native C/C++ binaries is not that different from a JIT warming up C#/Java code, especially since llvm+clang is so good at optimizing code.

Often, you might want to run servers. If they are JIT compiled, then during warm-up they will be slow(er), and later they will be fast(er). But usually you don't restart servers that often, so they will spend most of their time in "fast mode". It makes sense to me to measure what will be used in practice.

Of course, for command line tools such as grep, it is more realistic to measure the warmup phase as well, as that will reflect real usage.

What makes you think those "non-AOT C# (and java)" have a "warming up" time that isn't included in the run time measurement?