We create multiple CQs per query when join optimization is triggered. Normally queries keep their CQ alive throughout execution- however, we only keep one CQ around, which could result in issues if anything accesses, for example, QSN filters, which don't own their backing BSON (under the expectation that this is owned by BSON in an MatchExpression in a CQ).

We don't have a repro for this yet, but we should fix the issue regardless.