> I don't see a clear demarcation of the things that only "reasoning" can accomplish.

Producing beautiful, coherent, maintainable code. RL trained the models to produce correct code but boy oh boy is it ugly if it's 10k+ LOC