fix(statistics): accept finite NumPy scalar measurements

[harshitethic]

Summary

• normalize NumPy scalar values with .item() before numeric validation
• keep rejecting Python and NumPy booleans as measurements/weights
• add regression coverage for np.float32, np.int64, and np.bool_

Fixes #647.

[greptile-apps]

Confidence Score: 4/5

[Medium risk] Statistics module accepts NumPy scalar types.

The PR should not merge until the isolated utility import test works again.

Findings

1.  Isolated utility import fails ▶
2.  Finite longdouble values fail ▶

Summary

WeightedValue now accepts finite NumPy number scalars as measurements and weights while still rejecting booleans.

• Converts NumPy scalars before numeric validation.
• Adds examples for accepted numeric scalars and rejected NumPy booleans.

_{Reviews (1) · Last reviewed commit: "test(statistics): cover NumPy scalar mea..."}

[greptile-apps]

Isolated utility import fails

The existing utility test runs Python without site-packages and imports hflow.statistics. This new numpy import raises ModuleNotFoundError, so the test fails before its checks run. Keep that import working without site-packages, or change the test and the documented standard-library-only promise together.

[greptile-apps]

Finite longdouble values fail

np.longdouble(1).item() does not become a Python int or float. The next check therefore makes WeightedValue(np.longdouble(1), 1) raise ValueError, even though the measurement is finite. Callers using that NumPy scalar type still cannot use these summaries; accept finite real NumPy scalars without relying on .item() to return a built-in number.

[kstonekuan]

Thanks for this. #647 was assigned to chiruu12, and their fix in #649 has now merged, so I'm closing this one. Issues with an assignee are taken; everything else is fair game. If you want another, exercising HFlow against a real corpus such as Egocentric-10K on Hugging Face and reporting what breaks or is slow is high-value work right now.