In the retryable_writes_jscore_stepdown_passthrough, startParallelShell is called with the connection string of the underlying replica set, so it immediately calls connect to set the global db object. If there is a stepdown during connect, the test will fail, because the auto_retry_on_network_error.js override is not loaded yet.

This was also a problem for the main test shell, and it was fixed by starting the shell with --nodb and passing an eval string that calls connect after loading the auto_retry passthrough. startParallelShell should be overriden in auto_retry_on_network_error.js to do the same thing.