diff --git a/mypy/checker.py b/mypy/checker.py
index d13b927b28f2..1316eaee6209 100644
--- a/mypy/checker.py
+++ b/mypy/checker.py
@@ -3138,6 +3138,16 @@ class C(B, A[int]): ...  # this is unsafe because...
             return
         first = base1.names[name]
         second = base2.names[name]
+        if name == "__replace__" and first.plugin_generated and second.plugin_generated:
+            # Plugin-synthesized __replace__ methods (e.g. those added by the
+            # @dataclass plugin on Python 3.13+ to support copy.replace())
+            # return Self and are regenerated fresh for every concrete
+            # subclass, so they can safely differ across unrelated bases --
+            # same reasoning as __init__ and friends above. We only skip this
+            # for methods a plugin generated, not ones the user wrote by
+            # hand, so real incompatible __replace__ overrides are still
+            # caught.
+            return
         # Specify current_class explicitly as this function is called after leaving the class.
         first_type, _ = self.node_type_from_base(name, base1, ctx, current_class=ctx)
         second_type, _ = self.node_type_from_base(name, base2, ctx, current_class=ctx)
diff --git a/test-data/unit/check-dataclasses.test b/test-data/unit/check-dataclasses.test
index 54b3afadc8b3..28d824bae20d 100644
--- a/test-data/unit/check-dataclasses.test
+++ b/test-data/unit/check-dataclasses.test
@@ -2609,6 +2609,65 @@ class Y(X):
 [builtins fixtures/tuple.pyi]
 
 
+[case testDunderReplaceDoesNotBlockAdHocIntersectionNarrowing]
+# https://github.com/python/mypy/issues/21635
+# flags: --python-version 3.13
+from dataclasses import dataclass
+
+@dataclass
+class A: ...
+@dataclass
+class M: ...
+@dataclass
+class B(A): ...
+@dataclass
+class C(M, A): ...
+
+alist: list[type[A]] = [B, C]
+mlist: list[type[M]] = [cls for cls in alist if issubclass(cls, M)]
+reveal_type(mlist)  # N: Revealed type is "builtins.list[type[__main__.M]]"
+[builtins fixtures/isinstancelist.pyi]
+
+[case testDunderReplaceDoesNotBlockPlainIssubclassNarrowing]
+# https://github.com/python/mypy/issues/21635
+# Same underlying bug as testDunderReplaceDoesNotBlockAdHocIntersectionNarrowing,
+# but without a list/comprehension: plain issubclass() narrowing in an
+# ordinary if-statement hit the exact same false "impossible intersection"
+# and silently marked the branch as unreachable instead of narrowing (no
+# reveal_type note, and the type error below went unreported).
+# flags: --python-version 3.13
+from dataclasses import dataclass
+
+@dataclass
+class A: ...
+@dataclass
+class M: ...
+@dataclass
+class B(A): ...
+@dataclass
+class C(M, A): ...
+
+cls: type[A] = C
+if issubclass(cls, M):
+    reveal_type(cls)  # N: Revealed type is "type[__main__.<subclass of "__main__.A" and "__main__.M">]"
+    n: int = 'foo'  # E: Incompatible types in assignment (expression has type "str", variable has type "int")
+[builtins fixtures/isinstancelist.pyi]
+
+[case testDunderReplaceHandwrittenStillCheckedForCompatibility]
+# flags: --python-version 3.13
+class A:
+    def __replace__(self) -> "A":
+        return A()
+
+class M:
+    def __replace__(self) -> "M":
+        return M()
+
+class C(M, A):  # E: Definition of "__replace__" in base class "M" is incompatible with definition in base class "A"
+    pass
+[builtins fixtures/tuple.pyi]
+
+
 [case testFrozenWithFinal]
 from dataclasses import dataclass
 from typing import Final