The problem was that should have been using a longest common subsequence
algorithm for the "number of word boundary character matches" calculation. Our
old approach would fail for the following case:
Query: "caafoo"
Candidate1 : "acaaCaaFooGxx"
Candidate2 : "aCaafoog"
Candidate1 needs to win. This is now also a test case.
The point is that we want to prefer candidates that have the query characters
"earlier" in their text, e.g. "xxabcxxx" over "xxxxxabc" for "abc" query.
