Starting with data types before touching a model is the right order, and it is the step people skip because it feels like admin rather than work. Most of the modelling mistakes I have had to unpick came from treating an ordinal as continuous or an ID as a number - both run fine and produce something that looks like a result. One concept worth adding early to a foundations series is variance of an estimate, not just of a variable. Knowing that a metric computed on 200 samples moves several points between splits is what stops you from chasing improvements that are noise, and it is the single idea that transfers most directly from statistics into day-to-day ML work.