In collection_defragmentation.js, a splitVector command is being used to get a split point for a random chunk. Since this operation is best effort anyways, it is simply being sent to any node from the shard, but some of the concurrency suites use the network error overrides in network_error_and_txn_override to retry network errors which includes the response notwriteableprimary. We should make sure to get the primary of the shard so that the test can complete or remove the call to splitvector entirely.