Looks like there's still lots of test_typing failures if I apply this diff to your PR branch:

diff --git a/Lib/typing.py b/Lib/typing.py
index 8c874797d2..8caeb915d8 100644
--- a/Lib/typing.py
+++ b/Lib/typing.py
@@ -1733,7 +1733,7 @@ def _allow_reckless_class_checks(depth=3):
     The abc and functools modules indiscriminately call isinstance() and
     issubclass() on the whole MRO of a user class, which may contain protocols.
     """
-    return _caller(depth) in {'abc', 'functools', None}
+    return _caller(depth) in {'functools', None}

The trouble is that these __subclasscheck__ calls call into _ProtocolMeta.__subclasscheck__, which might raise:

But we don't want it to raise if the __subclasscheck__ call is coming from an __instancecheck__ call.

Ultimately, it does sort of feel like typing is the module that's "in the wrong" here. The abc module makes the assumption that issubclass() calls will never raise, providing both arguments are instances of type. That's true for the vast majority of classes, so I think it's a reasonable assumption to make. typing is the module that's doing a surprising/unusual thing here, so it seems reasonable to say that it's typing that should have to do the awkward workaround. It's unfortunate how fragile that makes some of the code in typing.py, but I think PEP-544's semantics are too well-established now to change the behaviour.

I feel like this change isn't worth the churn.