Fork adds evolutionary DB to Karpathy's autoresearch

Karpathy's autoresearch hit ~35K stars in days — this fork bets that the original's flat experiment log is its weakest link, and evolutionary search is the fix.

• –MAP-Elites maintains a diverse "island" population of solutions rather than treating all past runs as a flat list, preventing the agent from over-exploiting a local optimum
• –Strategy-guided sampling (exploit/explore/random) gives the agent tunable search behavior — critical when each trial costs real GPU-hours
• –The lineage tree visualizer adds interpretability, letting researchers see how the search evolved across generations
• –The open question is whether evolutionary overhead pays off at autoresearch's scale: unlike AlphaEvolve's cheap simulations, each trial here costs 5 minutes of GPU compute
• –Early traction is minimal (3 stars, 0 forks) but the base project's prominence means any meaningful result here will get attention fast