Describe the Bug
scipy.fft.rfftn and scipy.fft.irfftn are inferred as tuple[Dispatchable], although they return NumPy arrays at runtime. This produces false-positive NumPy overload, indexing, and mutation errors in otherwise valid code.
Minimal reproduction (inside a configured project that resolves SciPy):
import numpy as np
from scipy import fft
def transform(img: np.ndarray, spikes: np.ndarray) -> np.ndarray:
axes = tuple(range(spikes.shape[-1]))
spectrum = fft.rfftn(np.ascontiguousarray(img), axes=axes, workers=1)
reconstructed = fft.irfftn(
spectrum,
s=tuple(img.shape[axis] for axis in axes),
axes=axes,
workers=1,
)
return np.clip(reconstructed, 0.0, 1.0)Expected: spectrum and reconstructed are inferred as ndarray, and the function type-checks.
Actual:
revealed type: tuple[Dispatchable]
No matching overload found for numpy.clip called with arguments: (tuple[Dispatchable], float, float)
The same tuple[Dispatchable] type also makes valid spectrum indexing and in-place addition fail.
Environment:
- Pyrefly 1.2.0
- Python 3.10.16
- SciPy 1.15.3
- NumPy 2.2.6
- macOS 26.4.1 arm64
The relevant SciPy API returns an ndarray at runtime. A local cast("np.ndarray", ...) is sufficient as a workaround, but loses the checker’s actual return-type information.
Sandbox Link
Not available: the repro requires the SciPy package and its resolved stubs in a configured local project.
(Only applicable for extension issues) IDE Information
N/A
Describe the Bug
scipy.fft.rfftnandscipy.fft.irfftnare inferred astuple[Dispatchable], although they return NumPy arrays at runtime. This produces false-positive NumPy overload, indexing, and mutation errors in otherwise valid code.Minimal reproduction (inside a configured project that resolves SciPy):
Expected:
spectrumandreconstructedare inferred asndarray, and the function type-checks.Actual:
The same
tuple[Dispatchable]type also makes valid spectrum indexing and in-place addition fail.Environment:
The relevant SciPy API returns an ndarray at runtime. A local
cast("np.ndarray", ...)is sufficient as a workaround, but loses the checker’s actual return-type information.Sandbox Link
Not available: the repro requires the SciPy package and its resolved stubs in a configured local project.
(Only applicable for extension issues) IDE Information
N/A