Assume unannotated __new__ returns Self

[grievejia]

Summary:
This fixes some pretty bad behaviors when dealing with return type
inference on complex code that is really dynamically typed, but can
be approximated by normal static typing (specifically, the
networkx _dispatch class).

In general, getting bad return type inference on constructors can
fan out to a lot of errors, so this is a nice way to reduce FPs
against untyped dependencies that might do complicated, dynamic things.

It is consistent with the typing spec to do this. Note that networkx isn't
in mypy primer, but we actually get a significant error reduction on sympy which
is a good sign that this is a real improvement in practice on untyped code
beyond just the one motivating example.

Fixes #3120

Differential Revision: D100686394

[meta-codesync]

@grievejia has exported this pull request. If you are a Meta employee, you can view the originating Diff in D100686394.

[github-actions]

Diff from mypy_primer, showing the effect of this PR on open source code:

ppb-vector (https://github.com/ppb/ppb-vector)
- ERROR ppb_vector/__init__.py:244:16-58: Returned type `None` is not assignable to declared return type `Vector` [bad-return]
- ERROR ppb_vector/__init__.py:263:16-58: Returned type `None` is not assignable to declared return type `Vector` [bad-return]
- ERROR ppb_vector/__init__.py:292:16-56: Returned type `None` is not assignable to declared return type `Vector` [bad-return]
- ERROR ppb_vector/__init__.py:352:16-54: Returned type `None` is not assignable to declared return type `Vector` [bad-return]
- ERROR ppb_vector/__init__.py:463:17-30: `None` is not assignable to variable `other` with type `Sequence[SupportsFloat] | Vector | VectorLikeDict | tuple[SupportsFloat, SupportsFloat]` [bad-assignment]
- ERROR ppb_vector/__init__.py:467:13-25: Object of class `Sequence` has no attribute `length`
- Object of class `VectorLikeDict` has no attribute `length`
- Object of class `tuple` has no attribute `length` [missing-attribute]
- ERROR ppb_vector/__init__.py:468:28-44: Object of class `NoneType` has no attribute `length` [missing-attribute]
- ERROR ppb_vector/__init__.py:506:16-28: Returned type `None` is not assignable to declared return type `Vector` [bad-return]
- ERROR ppb_vector/__init__.py:583:17-30: `None` is not assignable to variable `basis` with type `Sequence[SupportsFloat] | Vector | VectorLikeDict | tuple[SupportsFloat, SupportsFloat]` [bad-assignment]
- ERROR ppb_vector/__init__.py:584:24-36: Object of class `Sequence` has no attribute `length`
- Object of class `VectorLikeDict` has no attribute `length`
- Object of class `tuple` has no attribute `length` [missing-attribute]
- ERROR ppb_vector/__init__.py:587:13-35: `*` is not supported between `float` and `Sequence[SupportsFloat]` [unsupported-operation]
- ERROR ppb_vector/__init__.py:587:13-35: `*` is not supported between `float` and `VectorLikeDict` [unsupported-operation]
- ERROR ppb_vector/__init__.py:587:13-35: `*` is not supported between `float` and `tuple[SupportsFloat, SupportsFloat]` [unsupported-operation]
- ERROR ppb_vector/__init__.py:588:16-27: Returned type `tuple[Vector | float, Vector]` is not assignable to declared return type `tuple[Vector, Vector]` [bad-return]
- ERROR ppb_vector/__init__.py:588:19-27: `-` is not supported between `Self@Vector` and `float` [unsupported-operation]
- ERROR ppb_vector/__init__.py:606:26-48: `None` is not assignable to variable `surface_normal` with type `Sequence[SupportsFloat] | Vector | VectorLikeDict | tuple[SupportsFloat, SupportsFloat]` [bad-assignment]
- ERROR ppb_vector/__init__.py:607:24-45: Object of class `Sequence` has no attribute `length`
- Object of class `VectorLikeDict` has no attribute `length`
- Object of class `tuple` has no attribute `length` [missing-attribute]
- ERROR ppb_vector/__init__.py:610:16-69: `-` is not supported between `Self@Vector` and `float` [unsupported-operation]
- ERROR ppb_vector/__init__.py:610:24-68: `*` is not supported between `float` and `Sequence[SupportsFloat]` [unsupported-operation]
- ERROR ppb_vector/__init__.py:610:24-68: `*` is not supported between `float` and `VectorLikeDict` [unsupported-operation]
- ERROR ppb_vector/__init__.py:610:24-68: `*` is not supported between `float` and `tuple[SupportsFloat, SupportsFloat]` [unsupported-operation]
- ERROR ppb_vector/__init__.py:617:17-44: Object of class `NoneType` has no attribute `normalize` [missing-attribute]
- ERROR ppb_vector/__init__.py:622:17-29: `None` is not assignable to attribute `x_unit` with type `Vector` [bad-assignment]
- ERROR ppb_vector/__init__.py:623:17-29: `None` is not assignable to attribute `y_unit` with type `Vector` [bad-assignment]
- ERROR ppb_vector/__init__.py:624:15-27: `None` is not assignable to attribute `zero` with type `Vector` [bad-assignment]
- ERROR tests/test_angle.py:24:10-20: Object of class `NoneType` has no attribute `angle` [missing-attribute]
- ERROR tests/test_angle.py:25:10-21: Object of class `NoneType` has no attribute `angle` [missing-attribute]
- ERROR tests/test_decompose.py:35:72-76: Argument `None` is not assignable to parameter `other` with type `Sequence[SupportsFloat] | Vector | VectorLikeDict | tuple[SupportsFloat, SupportsFloat]` in function `ppb_vector.Vector.isclose` [bad-argument-type]
- ERROR tests/test_decompose.py:36:72-76: Argument `None` is not assignable to parameter `other` with type `Sequence[SupportsFloat] | Vector | VectorLikeDict | tuple[SupportsFloat, SupportsFloat]` in function `ppb_vector.Vector.isclose` [bad-argument-type]
- ERROR tests/test_length.py:23:12-25: Object of class `NoneType` has no attribute `length` [missing-attribute]
- ERROR tests/test_member_access.py:10:12-15: Object of class `NoneType` has no attribute `x` [missing-attribute]
- ERROR tests/test_member_access.py:11:12-15: Object of class `NoneType` has no attribute `y` [missing-attribute]
- ERROR tests/test_pattern_matching.py:30:21-22: Cannot match positional sub-patterns in `Never` [bad-match]
- ERROR tests/test_pattern_matching.py:30:24-25: Cannot match positional sub-patterns in `Never` [bad-match]
- ERROR tests/test_project.py:30:88-92: Argument `None` is not assignable to parameter `other` with type `Sequence[SupportsFloat] | Vector | VectorLikeDict | tuple[SupportsFloat, SupportsFloat]` in function `ppb_vector.Vector.isclose` [bad-argument-type]
- ERROR tests/test_project.py:50:29-33: Argument `None` is not assignable to parameter `other` with type `Sequence[SupportsFloat] | Vector | VectorLikeDict | tuple[SupportsFloat, SupportsFloat]` in function `ppb_vector.Vector.isclose` [bad-argument-type]
- ERROR tests/test_reflect.py:22:12-35: Object of class `NoneType` has no attribute `reflect` [missing-attribute]
- ERROR tests/test_rotate.py:87:35-38: Object of class `NoneType` has no attribute `x` [missing-attribute]
- ERROR tests/test_rotate.py:87:41-44: Object of class `NoneType` has no attribute `y` [missing-attribute]
- ERROR tests/test_rotate.py:147:17-31: Object of class `NoneType` has no attribute `rotate` [missing-attribute]
- ERROR tests/test_rotate.py:152:20-35: Object of class `NoneType` has no attribute `rotate` [missing-attribute]
- ERROR tests/test_rotate.py:156:43-57: Object of class `NoneType` has no attribute `length` [missing-attribute]

[github-actions]

Primer Diff Classification

✅ 1 improvement(s) | 1 project(s) total | -35 errors

1 improvement(s) across ppb-vector.

Project	Verdict	Changes	Error Kinds	Root Cause
ppb-vector	✅ Improvement	-35	bad-argument-type, bad-assignment	is_function_without_return_annotation()

Detailed analysis

✅ Improvement (1)

ppb-vector (-35)

This is a clear improvement. The PR implements spec-compliant behavior per https://typing.readthedocs.io/en/latest/spec/constructors.html — unannotated __new__ methods should be assumed to return Self. The Vector class has an unannotated __new__ that correctly constructs and returns Vector instances. Previously, pyrefly inferred None as the return type, causing 35 cascading false positive errors across the project. All removed errors are false positives that arose from this single inference failure.
Attribution: The change in pyrefly/lib/alt/class/class_field.rs adds logic to detect when __new__ has no return type annotation (is_function_without_return_annotation()) and assumes it returns Self. This directly fixes the root cause: Vector.__new__ has no return annotation, so pyrefly now correctly infers it returns Vector (or the appropriate subclass via Self), eliminating all 35 cascading false positives.

---

Was this helpful? React with 👍 or 👎

_{Classification by primer-classifier (1 LLM)}