I agree with all points except the last — everyone should move to production already with the baseline model! This is crucial for the same points you make and also for actually evaluating how much business value the simplest model brings. It also helps with iteration, since every single new version of the model ends up straight in production where it can be evaluated on the business KPI that matters and not the proxy — in my experience it is very difficult to get a project where you can evaluate ML solution solely on test set.