[JonSnow1807]

Rationale for this change

There is no kernel for multiply(duration, float64) or divide(duration, float64), so these calls raise NotImplemented (see #48003).

What changes are included in this PR?

Kernels for multiply(duration, float64), multiply(float64, duration) and divide(duration, float64), registered for multiply, multiply_checked, divide and divide_checked. The result keeps the duration’s unit and is truncated toward zero, like numpy and divide(duration, int64). NaN, overflow and division by zero return an error. Only float64 is handled; float32 and float16 are not cast.

Are these changes tested?

Yes, a new test in scalar_temporal_test.cc (truncation, both argument orders, nulls, errors, all four time units) and a small Python test with the example from the issue.

Are there any user-facing changes?

Yes, these calls now return a duration instead of raising NotImplemented.

Was AI used for this PR?

In accordance with the AI generation guidelines, please disclose below whether and how AI was used in this PR.

PR code and description written by:

• Human
• AI

Reviewed before submission by:

• Human
• AI
• Not reviewed

Code drafted by both human and AI, and reviewed by human.

• GitHub Issue: [C++][Python] Add multiply/divide kernels for duration and float/double #48003

[github-actions]

⚠️ GitHub issue #48003 has been automatically assigned in GitHub to PR creator.

[rok]

Please do adversarial review with your agents. Have them look at other kernels. Once you've done that please review this yourself such that you know what each line does, then ping me.

[rok]

Also - add a small Python test that will use example proposed by Randolf in #48003.

[JonSnow1807]

Thanks for the guidance. I ran an adversarial review with two separate reviewers and compared the kernel against the existing ones (round, checked arithmetic, casts, divide(duration, int64)). From that, the kernel now uses the physical Int64Type like its neighbours, the test also covers the lower int64 bound, and the docs say the computation is done in float64. I switched to truncation toward zero, so Randolf’s example now gives exactly numpy’s output, and I added it as a Python test. I’ve gone through every line myself. Also, to correct my earlier comment: float32 and float16 are not cast in this PR.

[rok]

Thank you. I'll try to take a look tomorrow.
cc @randolf-scholz