So after testing Magistral and the Magistral-Cydonia merge (v3j) I've come to the conclusion that assistant reasoning ("Okay, so here's an RP scenario...") is useless. Obviously it's a bit more creative since it's trained on reasoning outputs (I keep seeing words typed in Chinese characters in Magistral's reasoning), but it's not that different.
However, everyone seems to be ignoring in-character reasoning. It's the default reasoning mode for these models (i.e. just <think>\n prefill). It lets the model state all the exposition slop, hidden motives, etc. in-character. This means all that slop won't be inserted in the actual reply.