action_l2. Anti-saturation against tanh saturation: relu(|a| − threshold)² × coeff.
portfolio_budget. Quadratic penalty for exceeding the gross-exposure limit.
inaction_penalty. Anti-zero-policy — pushes the actor away from zero positions.
action_diversity. Anti-symbol-collapse — penalty on low std(actions, dim=symbols).
direction_inbalance. Anti-mono-direction — penalty when |batch_mean(actions)| exceeds threshold.