The error indexes in userBatchIndices are already real indices and don't need start added to them again. This is very difficult to hit in production: it requires pre-existing data corruption in a bucket that is not the first bucket an insert writes to.