When adding global_random_seed to a test, it's not enough to check it locally, i.e. on a single machine. Numerical precision issues can come from various factors like OS, CPU, BLAS, ...

When adding global_random_seed, it's important to test all random seeds on all CI jobs. To do that, you need to push a commit with [all random seeds] and the list of tests to check in the commit message:

some message [all random seeds]
test_something
test_some_other_thing

NOTE: It is mandatory to pass a short list of test function names after the [all random seeds] commit flag. Running the full scikit-learn test suite for all random seeds at once would take too long.

If this is not done, we merge the PR and then the nightly builds fail every once in a while because the tolerance was barely too small for some seed.