Fix pairwise inference task index selection

[SergeiNikolenko]

Description

Fix pairwise inference dataframe generation to use explicit task index selection (task_idx) and add regression tests for single-task and multitask pairwise paths.

Status

• Ready to go

Developers certificate of origin

• I certify that this contribution is covered by the MIT License here and the Developer Certificate of Origin at https://developercertificate.org/.

[SergeiNikolenko]

Closing per author request during fork cleanup/reset.

[hmacdope]

@SergeiNikolenko does this have a specific task that is not working or issue that this is associated with? I wasn't aware that anything was wrong.

[SergeiNikolenko]

Hi, yes, this is fixing a real bug in the pairwise inference path.

Before this change, _generate_pairwise_df was using predictions[:, j], where j came from the generated pair (i, j) rather than from the task index. That means the selected prediction column depended on the pair indices instead of the output task.

In practice, that could lead to two problems:

1. for multitask pairwise inference, predictions could be written from the wrong output column;
2. for single-task pairwise inference, this could hit an out-of-bounds column index depending on the generated pairs.

I added tests covering both the task index selection and the pairwise inference path.

[smcolby]

The fix to inference.py is correct and the core test (test_generate_pairwise_df_uses_task_idx_column) is solid. A few changes are required before merge.

1. Test style: replace monkeypatch + hand-rolled dummies with pytest-mock

test_predict_single_task_pairwise_uses_column_zero and test_predict_pairwise_multitask_passes_task_idx both use monkeypatch.setattr with custom nested dummy classes (DummyPairwiseFeaturizer, DummyModel, fake_loader, fake_generate_pairwise_df). We're trying to use pytest-mock with mocker.patch(..., autospec=True) for this sort of thing, and will be updating the contribution guidelines soon.

Please rewrite both tests using the mocker fixture, e.g.:

def test_predict_single_task_pairwise_uses_column_zero(mocker, tmp_path):
    mock_load = mocker.patch(
        "openadmet.models.inference.inference.load_anvil_model_and_metadata",
        autospec=True,
    )
    mock_load.return_value = (...)
    ...

test_generate_pairwise_df_uses_task_idx_column is fine as-is — keep it.

2. DummyPairwiseFeaturizer is copy-pasted verbatim between the twomonkeypatch tests. Once rewritten with mocker, this duplication goes away naturally.

[dwwest]

This looks good, thank you for catching this bug. Just a quick cosmetic change, plus Sean's requests re: testing, and it should be good to go.

[dwwest]

Let's change j to task_idx for consistency.