PR summary

Closes #31210.

This PR fixes Generator.choice(..., replace=False, p=..., shuffle=...) so that shuffle is respected in the weighted sampling path.

Right now, when replace=False and p is provided, shuffle=True and shuffle=False produce the same ordering. That is different from the unweighted path, where shuffle does affect the output order. In practice, this means the draw order can reflect the internal weighted sampling order instead of being randomized, which is exactly the behavior reported in the issue.

The code change is small: after the weighted no-replacement sampler finishes building the selected indices, it now applies the same final shuffle step that is already used in the unweighted no-replacement path when shuffle=True.

I also added a regression test to check that:

• weighted sampling without replacement now changes order when shuffle=True
• the selected elements are still the same set
• the existing deterministic test for the non-shuffled weighted path continues to work by making it explicitly use shuffle=False

Tests run in Docker:

• spin test numpy/random/tests/test_generator_mt19937.py -- -q -k "test_choice_nonuniform_noreplace or test_choice_nonuniform_noreplace_shuffle or test_choice_uniform_noreplace"
• full Docker test run with:
  • spin build --clean -- -Dallow-noblas=true -Dcpu-baseline=none -Dcpu-dispatch=none
  • spin test -- --durations=10 --timeout=600

The focused random tests passed. The full test run completed except for the known environment-specific failure in numpy/_core/tests/test_cpu_features.py::Test_X86_Features::test_features.

I also reproduced the issue visually using the same kind of setup described in the issue report and confirmed that the fixed shuffle=True behavior removes the order correlation, while preserving the sampled set.

AI Disclosure

I used ChatGPT in preparing this pull request.

AI was used for grammar and wording fixes in PR text, and earlier I did not fully disclose that use. I’m sorry about that. It was not intentional.

For completeness: AI was also used to help inspect the code path and help draft text while I was working on the issue. The actual code changes, testing, and submission are being handled by me.