Oh, I see.. The temp ClassLoader should not be used (for now) to build the ClassLoaderAccess during MetadataBuildingProcess#complete. But what I had meant to do was to instead revoke the temp ClassLoader from the MetadataBuilder at the start of the second phase. Seems like I just forgot to do that part.

Would be (obviously) helpful to be able to add a regression test for this. I have been thinking about ways to test this. Any thoughts?