For reference this test is triple-underscore prefixed, meaning it's not necessarily expected to work.
Sounds like it is making some unneeded or undocumented assumptions though, which should be corrected.

I believe I changed it to triple-underscore some time after making this issue.

I changed the semantic meaning of underscores.
No = standard test, can run concurrently
Single = Slower test, can run concurrently but only in single-thread
Double = Slower test / messes with settings or could break other tests. Cannot run concurrently. Should be run sparingly.
Triple = Must be run via command line. Cannot run concurrently. Should be run sparingly.

Tests not expected to pass are indicated as such in comments inside the files.

FYI the above is no longer true...

We now have separate directories... first_tests, cli_tests, async_tests, and sync_tests.

The underscore semantics are as follows:
None = Expected to pass, and should be run on every patch, minor, and major release.
Single underscore = Expected to pass but may need special consideration. Should be run on every minor and major release, but not necessary for patch releases.
Double underscore = Not expected to pass and needs careful review. Should be run on every major release. Not necessary for minor or patch releases.