Progressive growth: brain grows a layer every 2nd dream (identity-init, params 348k→~908k), shown live

Fix mode-collapse: repetition penalty, typed praise=feedback (not vocab), seed-anchored stable training, gentler lr; newborn babbles before first dream

Add MoodNet (2nd tiny model): 40 affect-anchored moods drive speech + behavior; energy/dream loop

DREAM beat: deferred learning — babble while awake, fine-tune on dream

Drop unused numpy; requirements now just torch

Full pivot: from-scratch char-level baby that learns to talk + personality/evolution (CPU, off-grid)