If the source shard primary steps down during a migration, it can trigger one of the cleanupOnError scope guards in the MigrationSourceManager. This calls MigrationManager::_cleanup, which can call ShardServerCatalogCacheLoader::waitForCollectionFlush, which will uassert if the node is no longer primary, and because this was called inside a scope guard and isn't caught, the exception triggers std::terminate() and crashes the server.

Example failure: https://evergreen.mongodb.com/task/mongodb_mongo_master_enterprise_rhel_62_64_bit_coverage_concurrency_sharded_with_stepdowns_and_balancer_patch_e4ba7722773f68d42a66af7439e585cc2136d003_5a4ceaabe3c3316388000020_18_01_03_14_41_42/0