Reward formula mismatch: diversity bonus is not length-scaled in final reward

[cloak-one]

The implementation of adaptive reward appears inconsistent with the paper definition of final reward composition.
Paper formula:

$$ r_i = s_{L,i} \cdot (r_{\text{format},i} + r_{\text{div},i}) $$

Current code behavior effectively does:

$$ r_i = (r_{\text{format},i} \cdot s_{L,i}) + r_{\text{div},i} $$

So the diversity term is added after length scaling, instead of being scaled together with the format term.

Questions

1. Is this intentional? (i.e., should diversity bonuses be independent of response length?)
2. If not, should both components be scaled together: reward = (base_rewards_list[i] + diversity_bonuses[i]) * length_scalar?
3. Does the paper's formula in the submitted version match the actual intended implementation?