> 2^128 - 2^112 [...] it's significant, especially if you have a huge data center in Utah.
As an upper limit, assume that you remove 100*2^112. But that will still only eliminate 100/65536=0.1% of the search space. Any key that is brute-forceable by NSA with those 0.1% removed is also brute-forceable without those 0.1% of the search space removed.
> What may be worse (I don't know) is the simultaneous equations that it creates that are invariant for keys from such a source. Maybe they could be used in a cryptographic attack to help solve the sorts of attack that try to build big systems of simultaneous equations to attack the key schedule.
Something like this seems slightly more likely. But assuming the bits were perfectly random before the removal of repeated blocks, for finite keys it still doesn't generate anything that couldn't have been generated by chance without the removal of repeated blocks.