Never seen this sort or primary endpoint and analysis in an RCT

This is a great discussion. One of the most important benefits of the PO model (and other models in the cumulative probability model family) is that even when the parallelism assumption is grossly violated, the model still tells you which treatment is better from a stochastic pairwise ordering perspective. The standard error of the log odds ratio may be screwed up, but the sandwich estimator is likely to fix that. Better: put a prior on the amount of non-PO.