Right now we don't normalize impl bounds during method probing, which means that the winnow stage may not work great when those bounds involve associated types. I think with the new setup we should be able to do this but didn't have time to test. Search for the FIXME in probe.rs