Justify Choosing PPO Over TRPO for a Production RL System

Your team is building a production RL system to optimize bidding decisions in an ad auction, with a shared neural network trunk feeding both a policy head and a value head, trained with minibatch SGD on GPU infrastructure already used for the rest of the ML stack. A research-minded colleague argues you should implement TRPO instead of PPO because "it has a real theoretical guarantee, and PPO's clipping is just a heuristic."

1. Give a substantive response to the colleague: is their characterization of PPO as "just a heuristic" fair? What does PPO actually retain from the motivation behind TRPO's guarantee, and what does it give up?
2. List the concrete engineering reasons PPO fits this system's architecture and infrastructure better than TRPO.
3. Describe one realistic scenario where a team might still prefer TRPO (or a KL-constrained approach) despite these costs.