You seem to be saying that a computer can only be called a quantum computer if it never decoheres. Or perhaps that it is only usable in some sort of batch-mode, like mainframes with punch-card decks, with each calculation cycle being a laborious one-shot affair.
I have little doubt that many near-future quantum computers will be used in exactly that mode, but also that at some point thereafter their coherence/decoherence cycles will start to be measured in Hz, Khz, etc.
Therefore there should in principle be no barrier to a quantum computer participating in a Turing test, or any other use-case that requires interaction.
Not quite. I'm saying that in the context of using a quantum-computer to play the role of Wigner's friend, it must be isolated from the ultimate observer. If you're not going to require that, you might as well just do the experiment with an actual human.
An AI could use Grover search when trying to find a good enough response. That would be like part of the AI going into and out of superposition for an efficiency boost.
Many quantum tasks are more efficient if certain catalyst states are present [1][2]. That would be like a part of the AI staying permanently in superposition while still ultimately contributing to classical results, again for an efficiency boost.