Three-state machine (warmup → capture → replay) wraps the post-fill section: extract_realized_pnl_delta, unit_state_update, trade_context, multires, step_counter, 3× EMA, abs_copy, reward_shaping, raw_rewards snapshot, apply_reward_scale, reward_clamp, atom_support_update, compute_advantage_return, var_over_abs_mean. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>