I was thinking of recording the production requests for x amount of time and replaying them on HHVM so you could see if responses are identical, there are no crashes, state of the database is same afterwards, etc.?
Seriously though, I decided that if I had a project that was difficult to test, that I knew had hairy code that used some funky features it was best to just stick to APC. All my new projects are being built on HHVM directly with test suites as I go.
Edit; that was a while back now, though. What frameworks and libraries are you using? Run their test suites through HHVM and if they all pass, you'll be fine. Facebook use it themselves now apparently.
Another idea once you're reasonably comfortable and have done a smoke test is to direct a small amount of your traffic/users (say 1%) to the HHVM version, while closely monitoring its error log.