Issues such as T132645 and T205675 were confusingly reported to the error channel in Logstash. This made me wonder "How could execution continue after this?" and if it really did continue, then "Why is there not also a fatal error in the logs?" (given every line after the line reported the error from would spew a "Cannot call on non-object null" error, but nothing like that is found.
Turns out, they are actually fatal (as they should be), and execution did stop. But due to HHVM's weird and magical error handling feature that reports warnings and fatals through the same callback, it is currently ending up in our "error" channel.