A model is an artifact tied to three moving parts
A trained model isn't a single, self-contained thing, it's the output of three ingredients combined at a specific point in time: the training data, the feature engineering and preprocessing code, and the model training code and hyperparameters. Change any one of those and you get a different model, even if the algorithm and architecture stay identical. This is the same reason software engineers version their code, except here the 'code' includes the data too.
Without tracking all three together, you lose the ability to answer basic questions later: exactly what data was this model trained on? Which version of the feature pipeline produced its inputs? What code and settings generated it? If you can't answer those, you can't reproduce the model's behavior, debug why it's making a strange prediction, or explain its decisions to someone who needs to trust it.