The initialization phase of resharding samples some documents from the source collection in order to pick the split points for the initial chunks. This is done by sending an aggregation with $sample over the new shard key.

When the collection being resharded is timeseries, at this point the new shard key is already translated to the bucket fields (i.e. "raw") and the aggregation pipeline refers to those fields. However currently the outgoing aggregation command is not sent with rawData, so the aggregation is being executed over the measurements instead of the buckets as intended.

Send the aggregation command for that pipeline with rawData.