Multiple variations "per-card" are the way to go. It was one of my big gripes about Anki too! My other big complaint is that it only really is tracking one metric when it comes to study - whether you remember the "card" or not.

This is where Anki falls short for language study, because we as language learners also want to look at listening & reading comprehension as well as pronunciation.

Any reason why you're using SM-2 over FSRS? In Emurse we're using an algorithm similar to SM-2, mainly because FSRS wasn't out when we started building it. it works well enough because we've come up work arounds for Anki's problem of workload distribution. Looking into FSRS a bit, I am definitely interested in giving this a shot and seeing if it improves it in that area.