Perhaps. Nondestructive editing tends to preserve the original image and applies edits to a copy. It
could do that in a way that makes that possible, but it's likely someone would need to specifically build the capacity to remember text layers like that.
A naive implementation might just have two buffers and a list of pixel changesets and no knowledge of how the changes were made. A more advanced implementation could track more information, work on a per layer basis, and support the recovery of vector text layers from rastered ones.
So I guess the answer is not by default, but it opens the door for it to be a thing.
