Handling of np.int64/np.float64 in `Series.sum()`

[loicdiridollou]

Describe the bug

At runtime, Series.sum returns a np.int64 or np.float64 when summing numbers. Yet the typing says we should get an int or a float.
The issue arises when we use method on that result that are only defined on np.int64 or np.float64 (for example the clip method).

How would we want to handle those cases, are we open to adjusting the stubs to better type the behavior?

To Reproduce

1. Provide a minimal runnable pandas example that is not properly checked by the stubs.
2. Indicate which type checker you are using (mypy or pyright).
3. Show the error message received from that type checker while checking your example.

import pandas as pd
import numpy as np


a = pd.Series([1.,1.])
b = pd.Series([2.,2.])

a.mul(b).sum().clip(max=3)

The error we get is Cannot access attribute "clip" for class "float"

Please complete the following information:

• OS: MacOS
• OS Version latest
• python version 3.13
• version of type checker 1.18
• version of installed pandas-stubs main

Additional context
Add any other context about the problem here.

[cmp0xff]

The result depends on the backend of the Series.

>>> import pandas as pd
>>> pd.Series([1., 2.]).sum()  # backend: np.ndarray, result: numpy floating
np.float64(3.0)
>>> pd.Series([1., 2.], dtype="Float64").sum()  # backend: pd.FloatingArray, result: numpy floating
np.float64(3.0)
>>> pd.Series([1., 2.], dtype="float64[pyarrow]").sum()  # backend: ArrowExtensionArray, result: builtin float
3.0

In #1110 (comment) I am proposing adding the backend as a new generic variable to Series and Indexes. I have a few work-in-progresses in my fork, but they are very hard to work out.

[cmp0xff]

I think 2/3 of the backends use numpy types. For now maybe we can redirect the element types and the results of such functions to the numpy types.

>>> pd.Series([1., 2.]).iloc[0]
np.float64(1.0)
>>> pd.Series([1., 2.], dtype="Float64").iloc[0]
np.float64(1.0)
>>> pd.Series([1., 2.], dtype="float64[pyarrow]").iloc[0]
1.0

[Dr-Irv]

I think 2/3 of the backends use numpy types. For now maybe we can redirect the element types and the results of such functions to the numpy types.

We can try that, but I think it will open up a whole host of issues. I don't necessarily think that users should expect numpy types from the results of pandas operations.

The docs are unclear here because the examples show a result of a float, not numpy float. I think this has to do with how numpy changed the __repr__() or __str__() of its results at some point.

I created a docs issue for discussion in the pandas team: pandas-dev/pandas#62966