[toast0]

I'd expect all caching recursive resolvers to cache glue records as well as the requested records. It's part of the deal; if they didn't, the root servers would have an insane load. (After all, you need to get glue records for .com, etc, too)

Of course, there's probably exceptions, and cache size is limited. I can say that some resolvers seem to cache glue records for a long time, more than the published TTL; when I switched NS records for a popular domain, we continued to get requests on the old domain for more than 4 weeks (at which point we needed to end the previous service).