Traditional replay ------------------ turn 1: system + user 1 turn 2: system + user 1 + assistant 1 + user 2 turn 3: system + user 1 + assistant 1 + user 2 + assistant 2 + user 3 Stateful inference ------------------ wake: prefill stable system and continuity context once turn 1: append user 1; retain generated sequence and KV turn 2: append user 2; retain generated sequence and KV turn 3: append user 3; rebuild only when the causal prefix is no longer valid