Task 1: Allocate aux streams + workspaces + events Task 2: IQN/Attention accept workspace+stream override Task 3: Fork-join in submit_aux_ops (IQN+Attention parallel) Task 4: Switch capture mode GLOBAL → RELAXED Task 5: Forward backward branch parallelism Task 6: Fix adam child event timing (verify) Task 7: Deploy + verify child graph breakdown Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>