Hey, HHVM alum here.
The outright language gaps to 5.4 are few and fairly obscure these days, in my opinion. Most practical problems are due to missing or slightly misbehaving extension functions, rather than the core language. You can write programs that produce different output from php.net if, e.g., they depend on the order that destructors are called when killing lots of objects at once, but you can also write programs that differ across minor revisions of php.net. In the absence of a standard, much like pre-standardization C, "PHP" is anything that runs useful PHP programs. We think HHVM qualifies.
HHVM lets you do eval(), and we try to do a decent job of running it fast. We do actually jit eval'ed code, though if you're running in whole-program optimization mode the presence of eval() will prevent lots of compile-time class hierarchy analysis based optimizations from working. Yes, yes, we should be optimistic and deoptimize when we're wrong, it's just easier said than done :).
Facebook disallows eval() in production code for reasons you can probably imagine.