The race condition can be reproduced on Linux using a lot of parallel test worker processes. For example, I reproduce it on my laptop (8 CPUs) using:

$ ./python -m test test__xxsubinterpreters -F -j30
...
0:00:26 load avg: 17.86 [ 24] test__xxsubinterpreters passed
0:00:26 load avg: 17.86 [ 25] test__xxsubinterpreters passed
0:00:26 load avg: 17.86 [ 26] test__xxsubinterpreters passed
0:00:27 load avg: 17.86 [ 27/1] test__xxsubinterpreters failed
Exception in thread Thread-8:
Traceback (most recent call last):
  File "/home/vstinner/python/master/Lib/threading.py", line 944, in _bootstrap_inner
    self.run()
  File "/home/vstinner/python/master/Lib/threading.py", line 882, in run
    self._target(*self._args, **self._kwargs)
  File "/home/vstinner/python/master/Lib/test/test__xxsubinterpreters.py", line 51, in run
    interpreters.run_string(interp, dedent(f"""
RuntimeError: unrecognized interpreter ID 46
test test__xxsubinterpreters failed -- Traceback (most recent call last):
  File "/home/vstinner/python/master/Lib/test/test__xxsubinterpreters.py", line 492, in test_subinterpreter
    self.assertTrue(interpreters.is_running(interp))
AssertionError: False is not true