Updated title + more context from max.hirschhorn@mongodb.com's comment

While the manifestation a --dbg=on build variant would be the for mongos process to crash, in a production system, a client which had specified {readConcern: {level: "snapshot", atClusterTime: T1}} as the transaction read concern would have used a different timestamp T2 > T1 when starting the transaction on later participants. This would result in different participants using different read timestamps for the cross-shard transaction and violate snapshot isolation.