docs(readme): audit fixes + Llama 1B sealed eval + 3-architecture results

feat(eval): add xkcd / hand-drawn style option to reward curve plot

feat(eval): combined two-architecture reward curve plot (rubric headline)