That's an interesting idea but would set up a mismatch between the text representation and other representations. In most representations, for example JSON, we want to be able to process just the body of the response to get the "meaning" of the representation. If we have that pattern established for JSON and other data-based representations, having a different one for text would be confusing and conflicting.

Also, putting the key-value stuff in the headers would require a lot intelligence about naming the keys and processing the values before sticking them in the headers so that they are HTTP compliant.