On the donor shard, the MigrationChunkClonerSourceLegacy::startClone code uses _callRecipient private class function to call the recipient, which then uses the task executor to make the call. The task executor does not retry NotMaster errors.

A solution would be to use a ShardRemote, instead, and allow NotMaster errors to be retried for that first command, _recvChunkStart – don't want to use it for all commands, but the first one is safe, I think.