1. The comparison for matching terms used only the CRC to decide if there's
a match. Two different terms with the same CRC gave a match.

2. It assumed that if the second operand has more terms than the first, it's
never a match. That assumption is bogus, because there can be duplicate
terms in either operand.

Rewrite the implementation in a way that doesn't have those bugs.

Backpatch to all supported versions.