"Just shift back" is really underestimating how much effort it takes to port a design to a different foundry. Sure, you can target a new stdcell library and recompile your RTL (and re-floorplan, and re-do a bunch of other stuff) but you also have to swap out all your memories and interfaces, not all of which may have exact equivalents... it can easily take 1+ years of work for a competent team, and if you have to shift back all that time and effort was wasted.

Seems most major players are not only fabless but also fab-agnostic - as I noted they switch from one supplier to another even for the same product line. I'm sure it is work but it doesn't seem to be an existential crisis for a huge provider to send some volume to a new fab - certainly if it's derisking supply capacity, tariffs or other geopolitical risks.