> see the bug fix

Possibly re-tuning of some LLM parameters? Or forgetting some bad learnings... sounds like it's specific to a small-ish percent of users.