Measured data (192 middlegame roots from master games, 12-ply horizon). The raw policy castles in 0.0% of continuations where humans castle in 22.0% — yet at the very positions where humans castled, the policy's own value head ranks the castling move #1 in 43% of cases. The prior just starves it: median prior mass 0.0092, so search discovers the branch only in proportion to budget. This is the same starve-the-search wiring as KataGo's urgency horizon (post 4) with one decisive difference — there the value head shared the blindness (defense in depth, uncurable by budget); here it dissents, so every simulation added claws back behavior the prior suppressed.