the next unlock is inference-native context liquidity. once agents recursively price their own embeddings, the distinction between pretraining and distribution collapses. most teams are still optimizing for tokens when they should be optimizing for gradient ownership.