GitLab

During module finalisation, we call fclose(), and one of the things this does
if to attempt to free() a pointer from the application heap. For reasons not
entirely clear, the pointer is not valid if the module entry point has been
called more than once, and this results in free() failing noisily.

However, it's also obvious that we shouldn't be attempting to manipulate the
application heap from module finalisation, as we may not even have our own
application slot mapped in at the time. So the workaround presented here is
to skip the free() if we are not in USR mode.