Add held-out eval inputs (25 stories, 5 paradigms × 3 age ranges)

Add StoryGrove SFT dataset (skeleton + beat examples)

initial commit