> Yet... the first argument is still unshielded, passed to a macro that expects one single macro argument.

Are you talking about the current code in the master branch or the 3.9 branch? If you are talking about the _PyObject_CAST(ob) call, would you mind to elaborate how it is an issue? The code in master LGTM.

> That’s not a regression, it wasn’t shielded in 3.8 either, but why not just parenthesise each macro argument that denotes an expression (as opposed to e.g. name)?

In the master branch, Include/ and its subdirectories contains 158 .h files for a total of 20K lines of C code. It's not easy to check all files. If you spotted bugs, would you mind to give examples, or a way to automatically detect bugs?

As I wrote previously, I dislike macros. If someone is changed, I would prefer to convert the function into a static inline function which doesn't have macros pitfalls.