The keys aren't going into a single bucket above. The Vars() is just stripping out the non-parameter keys that CGI.pm stores inside of itself. The 'a', 'b' parameters are there in the object.
The way to determine what is happening inside a hash is evaluating it in scalar context. That gives you the number of buckets being used. tilly wrote a program that uses this feature to generate a list of colliding keys. This algorithm is fast and doesn't depend on reverse engineering the Perl hash algorithm.
I ran some tests on a 10,000 keys generated by tilly's method. Both inserting them into a hash and parsing the query string with CGI. It takes over 20 seconds to parse the query string in the pathological case versus less than a second for 10,000 normal strings. I haven't been willing to wait long enough to let 100,000 strings run. For a sample, here are the first 10 integers that collide and the scalar hash value showing they all go in one bucket.