Performance ratings as a control variable when the ratings may themselves be skewed?

Your own data suggests the answer could differ by function. You could include rating for research, where the distributions match, and exclude it for commercial. It is messier to explain, but it follows the evidence, and it tells the commercial leadership directly that their calibration is the issue.

I would not split the model by function for that purpose. You end up with two methods and an invitation to ask why you chose each. Report 2.5% as the headline, show that 1.1 points of it runs through ratings, and make the rating disparity in commercial a finding in its own right with its own action. One number, clearly decomposed.

Leaning towards the last suggestion, with the pre-test adopted as a standing rule for future years. Reporting the larger figure will be uncomfortable, but showing that 1.1 points of it passes through commercial ratings gives that leadership team something specific to own, which the blended 1.4% never did.