Looking good so far!

Toplevel comments:

1. I am not convinced yet that this indexing solution is correct yet. To help convince me, you can write a unittest that generates two large sets of random hashes (PDQSignal.get_random_signal()) b and q (e.g. 1,000 and 10,000), uses the brute force method to check exactly how many of hashes in b should match the hashes in q, then loads the b hashes into your indexing solution, and then queries all of the q hashes and makes sure you exactly get the expected results.
2. Many of your existing tests are tautological or should check to make sure the matching hashes are actually the ones that should be matching. You can likely replace many of your existing tests with variations on 2 above - I put some ideas for you in the IndexRecord class, but there are others that might work.
3. I recommend adding a method that can return a random hash exactly X bits different from a given hash, and then using that to help in your tests.

Nice work! We're nearly there, only a few more blocking comments!

So close!

1. The way you are seeding your hashes is causing duplicates in your query set - see inline blocking comment
2. As part of your test plan (doesn't have to be part of the test itself), please try generate a very large number of hashes and queries - e.g. 10,000 banked and 100,000 queries - something that takes 5-10 seconds to check. We want to be certain that the index is correct!
3. Please add a test that serializes and deserializes a non-empty index - you can do something cheeky like:

original = ... # do stuff to prepare index
deserialized = PDQIndex2,deserialize(index.serialize(...))
for index in (original, deserialized):
   # The original test goes here, should get same results on original or deserialized

All blocking comments resolved!