Sorry, if I was hand-wavy. I was trying to give other people a simple framework that I use myself (Big O calculated as a number of rows in each table). ALthough I dont know how many rows you have in each table, Big O frameworks still works, because cartesian of tables is being dominated by two huge different datasets (customers's orders being joined to employees' orders). The Union simply calcualtes them separately, rather than doing cartesian of two completely different datasets that represent different entities.

Well written predicates and indexes can help, as well as poorly written predicates make it worse. So there is balance. This is not a shortcut or a silver bullet, it is a trade-off being made. More indexes->faster selects and slower updates/inserts. One bad index=>failed insert and possible losing customer data (happened to me once)

I agree with you that "SQL performance is a complex topic." and one should definitely study query Execution Plan to understand the bottlenecks and make optimization decisions