“A U-shaped scaling law governs the trade-off between neural computation (MoE) and static memory (Engram), guiding optimal capacity allocation.”