I understand your concerns but feel that you may be forgetting many very useful use cases and, if I understand well what you mean by "the worst effect will be from developers who will widely practice the "easiest" way to achieve desired effect with dynamic xpath", I would think that any function, instruction, or feature can be misused and typically is, sometimes. That does not mean that we should remove them all, does it? If you built smart XML layouts, for example, with embedded rules, to query/fetch content, for example, would you define your own rule language and create your custom rule language interpreter, or would you not do that and rather hardwire all your queries in the applications, forcing users to modify the applications if they need to change/add/remove queries and/or layouts, or would you prefer using xpath and evaluate(), or do you have some other way of handling this?

I understand that there will be use cases almost for any feature, whatever.
But Occam's razor principle suggests to restrict oneself, otherwise you will
never know what do you really need, and what you don't.