I think what's missing is how to assess whether what you've build actually does what it's meant to do. Too many textbooks (actually almost all of academia) boil evaluation down to the calculation of a handful of narrow metrics.

But really understanding the failure modes of what you've made is much harder than that. Unfortunately I don't have a magic bullet or even a reference for that side of things. It's just something I've learned on the job. To paraphrase the Anna Karenina principle [0]:

All good models are alike; each bad model is bad in its own way.

[0] https://en.wikipedia.org/wiki/Anna_Karenina_principle