The chaotic mess part is not intended. Hopefully that won't happen now.
I've changed the twists prompting to hopefully prevent infeasible twists--though some of what you've mentioned is actually okay: enhancing a spell can work fine; it won't change the spell description but the narrator can generally handle it. Male side characters can be added as long as they aren't central to the story. Additionally, I've adjusted the twist mechanic to give you a choice of 3 twists, which should hopefully let you avoid any twists that break or undermine your run.
Going back to Command-R+ isn't really feasible as the API is still broken (4096 input tokens max for structured generations) whereas Command-A works with long contexts, allowing for much better narrator memory. Additionally, Command-A performs better along a lot of metrics, and a lot of the slop issues were very present in R+ as well. I actually do have some anti-slop built into the engine, but because I'm working through the API rather than a local instance of the LLM I can't use the most efficient measures (constraining the generation itself). I am adding a little more anti-slop though.