mypy resolves symbol incorrectly with builtins import cycle

Context: https://github.com/python/typeshed/pull/6289

To reproduce:

1. check out typeshed
2. git checkout 9eabedca5f47a710f29d6e3e2b3a389853fab5aa
3. apply this patch

diff --git a/stdlib/builtins.pyi b/stdlib/builtins.pyi
index 2393f3c9..0d2ea882 100644
--- a/stdlib/builtins.pyi
+++ b/stdlib/builtins.pyi
@@ -717,6 +717,7 @@ class bool(int):
     def __rxor__(self, __x: int) -> int: ...
     def __getnewargs__(self) -> tuple[int]: ...
 
+@final
 class slice(object):
     start: Any
     step: Any

Then run:

mypy --python-version 3.10 --no-incremental --custom-typeshed-dir . stdlib/_ast.pyi stdlib/ast.pyi 

and note that mypy incorrectly resolves slice to builtins.slice instead of ast.slice in ast.pyi

Note that if you change the order of files, it works just fine:

mypy --python-version 3.10 --no-incremental --custom-typeshed-dir . stdlib/ast.pyi stdlib/_ast.pyi 

It's a pretty weird case, import cycle with builtins, collision with name in builtins, from _ast import * in ast.pyi, etc. But a scary kind of bug :-)

9eabedca5f47a710f29d6e3e2b3a389853fab5aa のコミットにある stdlib/builtins.pyi、stdlib/_ast.pyi、stdlib/ast.pyi を使って再現から始めます。2 つのファイル順序で両方の mypy コマンドを実行し、ast.pyi が slice を builtins.slice に解決する理由を追跡します。完了条件は、報告されている順序依存の解決が、ファイル順序に依存せずに修正されていることです。