Aren't the false positives in this situation representative of set elements that node b thinks that node a has but really does not?

Right, the solution I remembered was a bit more complex, it required counting bloom filters and a subtraction operation to get a filter for the set difference. But the paper mentioned in a sibling comment claims to have even less overhead.