This is an investigative branch, not meant to be merged as such, to investigate what effort would be required to tackle #30007 as a whole (using pytest-run-parallel instead of pytest-freethreaded).

It includes commits from:

• MAINT replace enable_slep006 fixture by @config_context(enable_metadata_routing=True) #30038
• MAINT remove side-effects in test_partial_dependence #30039

That can be merged independently.

On top of this, I started to use @pytest.mark.parallel_threads(1) (EDIT: now @pytest.mark.thread_unsafe) on tests that are fundamentally not thread-safe or that use other fixtures such as tmpdir that would require to be initialized in the thread running the test to function properly (as opposed to running in the main pytest thread).

Many more similar changes to silence all the non-informative failures, but I will stop there for now.

This already highlights that the most common patterns are:

• testing for warnings with pytest.warns;
• testing for sys.stdout with capsys;
• usage of the tmpdir fixture;
• usage of the monkeypatch fixture.

EDIT: since the first version of this PR, pytest-run-parallel has been updated to make the tmpdir and tmp_path fixture work by default and automatically detect tests that use problematic fixtures and code patterns involving warnings automatically. The remaining problematic fixtures can probably be added by configuration.

I have also found a failure in test_minibatch_kmeans_partial_fit_init with the lambda init case using:

pytest --parallel-threads=4  sklearn/cluster/tests/test_k_means.py -k test_minibatch_kmeans_partial_fit_init

and I cannot explain it yet (using regular Python with GIL enabled).