When should a SQLAlchemy API query project columns instead of loading entities?

A read model is also a contract boundary. It prevents a serializer from walking persistence relationships and makes nullability or enum conversion deliberate. I align hot projections with indexes only after measuring the actual query plan, since covering every response creates write-heavy indexes. Tests verify mapping and pagination at the real database boundary, while telemetry tracks rows read, rows returned, payload size, and query duration.