I had to `TRUNCATE` the problem table so I don't have a perfect repro to completely convincingly test this anymore. Both queries behave fine at small scales, which is why we haven't seen this before.
I was able to run the newer query in production before I nuked the table and have it complete in a reasonable amount of time, while the old query hung longer than I wanted to wait (several minutes?). The query plan for the new query was also a good one, while the query plan for the old query was terrible.
I loaded the daemon console and ran `bin/garbage collect --collector worker.tasks --trace`. I verified the queries looked reasonable and produced reasonable results in production.