If it's important to you to remove the code creating a fake AST node here, you could instead add an extra argument to codegen_call_helper that is usually NULL and maps to an extra argument to add at the end of the list. I think I did that initially while working on PEP 695, but it seemed worse than creating a new virtual asdl_seq as the current code does. (Possibly you were involved in discussing that in person, don't remember exactly.)

But I'd like to keep the extra base class as an implementation detail in the compiler, not something that gets exposed in the AST layer.

Ok, I'll try something like that, should be quite simple.
It's possible I was involved in the discussion at the time, I don't remember the details.

I think it would be better if the compiler didn't need to know how to create AST nodes. (I take your point about it not modifying the AST, I'll correct the wording in the issue)

I found back the discussion from last time: #103764 (comment).