Fix crash with tuple unpack inside TypeVar default

[cdce8p]

Fixes #20451

[sterliakov]

IIRC the idea was that it should never be true when we call tuple_fallback - we have several places where things like tuple[*tuple[X,Y]] are normalized to a plain tuple, and the original intent was to get rid of such constructs as early as possible. This is fine as a quickfix, but I'm afraid that this will manifest again when such nested tuple is used elsewhere.

[cdce8p]

Not entirely sure about that but isn't the function already doing the flatting? E.g. just the line above flattens builtins.tuple Instances.

[sterliakov]

builtins.tuple is already normalized: you can't rewrite e.g. tuple[str, *tuple[int, ...]] as a single tuple[...] construct. This is only about expanding unpacks of fixed-size TupleType types (which is always possible), see flatten_nested_tuples

[cdce8p]

Think I found the right place now. The normalization happens in the TypeArgumentAnalyzer.

mypy/mypy/semanal_typeargs.py

Lines 110 to 115 in b69309b

	def visit_tuple_type(self, t: TupleType) -> None:
	t.items = flatten_nested_tuples(t.items)
	# We could also normalize Tuple[*tuple[X, ...]] -> tuple[X, ...] like in
	# expand_type() but we can't do this here since it is not a translator visitor,
	# and we need to return an Instance instead of TupleType.
	super().visit_tuple_type(t)

The visitor wasn't traversing TypeVar defaults.

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅