How would the analysis work if entities are accessed reflectively, e.g. by a mapping framework which creates XML structures representing a selected entity?

Good point. I think you're nailing the hardest aspect: I want this to be safe, hence even a first simple version would need to be able to recognize such a scenario and disable the optimisations. When unknown, or hard to understand, disable it. Smarter versions might then build on top of that potentially recognizing specific cases in which we can proceed with some optimizations.