So, I was finally able to replicate a failure in test_still_running locally, it required using a rather ridiculous number of parallel workers:

$ ./python -m test test__xxsubinterpreters -j200 -F
...
Exception in thread Thread-7:
Traceback (most recent call last):
  File "/home/aeros/repos/aeros-cpython/Lib/threading.py", line 944, in _bootstrap_inner
    self.run()
  File "/home/aeros/repos/aeros-cpython/Lib/threading.py", line 882, in run
    self._target(*self._args, **self._kwargs)
  File "/home/aeros/repos/aeros-cpython/Lib/test/test__xxsubinterpreters.py", line 51, in run
    interpreters.run_string(interp, dedent(f"""
RuntimeError: unrecognized interpreter ID 39
test test__xxsubinterpreters failed -- Traceback (most recent call last):
  File "/home/aeros/repos/aeros-cpython/Lib/test/test__xxsubinterpreters.py", line 766, in test_still_running
    interpreters.destroy(interp)
AssertionError: RuntimeError not raised
...
== Tests result: FAILURE ==

94 tests OK.

1 test failed:
    test__xxsubinterpreters

Total duration: 1 min 49 sec
Tests result: FAILURE

OS: Arch Linux x86_64
Kernel: 5.3.11
CPU: Intel i5-4460

I was able to consistently reproduce the above failure using 200 parallel workers, even without `-f`. 

I have a few different theories that I'd like to test for fixing the failure, I'll report back if any of them yield positive results.