I see there's an issue in the tracker to get more accurate data, and since it's using an under dev feature in compilers, it's not going to be definitive, but any rough numbers?
One reason is that, like you said, the feature is still new. Additionally I made pretty liberal use of the std::ranges library in my implementation which has notoriously bad compile times. That could be an area to improve upon.
Another may be a bit more structural. If you want to call myObj.foo() via reflection, you have to linearly search members_of(myObj) for the one named "foo", and then call that. Actual compilers I assume use some kind of hash table.
The hand-waving solution is "put it in a PCH", but I am hoping to put some more effort into optimizing build time here in the future.