What steps will reproduce the problem?
1. CityHash128() all 144-byte strings with 3 bits set
2. check for collisions

What is the expected output? What do you see instead?

0 collisions are expected.  87 collisions are actually seen.  I saw the first 
collision after about 8 million of the keys.  That's the expected collision 
rate for a 46-bit hash, not a 128-bit hash.

(I had a hash of my own last year, AkronHash, that was also a 128-bit hash 
evolved from MurmurHash.  I quickly withdrew it due to similar quality issues.  
My current attempt, SpookyHash, passes this test.)