The specific improvement scrypt makes over bcrypt is not yet relevant; nobody has ever hardware-optimized a bcrypt cracker, and the project that successfully does so and publishes their results will have made a contribution to cryptography literature.
Sure they have. A hardware-optimized bcrypt cracker is called a GPU. I can buy a 480-core GPU on Newegg for $350, but it doesn't come with any more RAM than a low-end PC does.
I am looking right now at Schneier's reference implementation of the Blowfish key setup function, and there is nothing in it that would not be straightforward to parallelize across a CUDA warp (read: no conditional branches). Is it your assertion that getting from here to a working bcrypt cracker would be sufficiently difficult as to constitute a significant "contribution to cryptography literature"?
Sure they have. A hardware-optimized bcrypt cracker is called a GPU. I can buy a 480-core GPU on Newegg for $350, but it doesn't come with any more RAM than a low-end PC does.