We currently store an encoded form of the final transformed version of each downloaded changeset in history during history integration. That means that for large changesets - especially the ones sent during initial sync that are 8MB each, we can end up holding 4 copies of each changeset and writing each changeset to disk at least twice as part of download integration. This can lead to significant memory pressure and write amplification.

This is not necessary, however. Once the downloaded changeset has been transformed and applied, we do not need to reference it again. Instead of storing the encoded transformed changesets, we should discard them and only update the progress counters contained in the download message.