You've probably thought more about this topic than I have, but it seems to me that words can at least be approximated as abstract referential entities, instead of just seen as a means of causing neuron stimulation in others. Using Putnam's proposed theory of meaning, I can build a robot that would bring me a biological-cat when I say "please bring me a cat", and bring the twin-Earth me a robot-cat when he says "please bring me a cat", without having to make the robot simulate a human's neural response to the acoustic vibration "cat". That seems enough to put Putnam outside the category of "dead wrong", as opposed to, perhaps, "claiming too much"?
I may bit a bit in over my head here, but I also don't see a strong distinction between saying "Assume on Twin Earth that water is XYZ" and saying "Omega creates a world where..." Isn't the point of a thought experiment to run with the hypothetical and trace out its implications? Yes, care must be taken not to over-infer from the result of that to a real system that may not match it, but how is this news? I seem to recall some folks (m'self included) finding that squicky with regard to "Torture vs Dust Specks" -- if you stop r...
Thagard (2012) contains a nicely compact passage on thought experiments: