I would use sad -> happy, because when users read it from left-to-right*, they will finish on the happy face. To the user, it subconsciously tells a story: "First this hypothetical person was sad, but then they became happy, therefore that makes me happy."
It's extremely subtle, and might not make a huge difference in the grand scheme of things. I would encourage you to do some A/B testing with both options and see if it has any effect on user behavior.
*Some languages are read right-to-left, so if you are planning to support any RTL languages, I would also test with natives to that language in case it makes a difference.