I feel similarly. When I've done anything structural it's been because I wasn't getting consistent results across similarly plausible specifications.

I almost entirely do reduced form estimation and that's what I primarily trained in. If you have a neat toy model of stacker behavior, I'd say go for it, but the incentives (and how well they're understood) don't strike me as being straightforward to model. It seems to me like you'd need to define different types of stackers and articulate the rewards system.

I would be very interested in trying to understand how closely behavior aligns with the rewards incentives, though. I've always thought that most of us are probably leaving lots of sats on the table by not optimizing.