This was discovered while running:

pytest -v --parallel-threads=4 --iterations=2 sklearn/svm/tests/test_sparse.py

before including the fix pushed to #30041 under bce2b4e.

I suspect the problem is that the built-in Platt scaling implementation of the vendored C++ code base of libsvm that uses a singleton pseudo random generator. Therefore, seeding the shared RNG state from competing threads prevents getting reproducible results and hence the test failure.