The guide
Most of the work in a real machine learning system is not the model. It is the data going in, and the model going out. Small mistakes here cost more than a poor choice of model, and they rarely raise an error.
The first three problems are about data. Missing values must be filled with statistics from the training set only. Splits must keep rare classes in the test set. Categorical features need an encoding that does not leak the label. Each one is a common source of results that look good in testing and fail in use.
The last three are about the model in production. Data drifts away from what the model was trained on, and PSI and the KS statistic are two common alarms for it. Quantisation makes the model small and fast enough to serve.