If the master fails over and a framework tries to do an explicit reconciliation for a task running on an agent that has not reregistered yet (and agent_reregister_timeout has not been exceeded), the master will not send a reconciliation response for that task.
This is confusing for framework authors. It seems better for the master to announce all the information it has explicitly: e.g., to return "task X is in an unknown state", rather than not returning anything. Then as more information arrives (e.g., agent reregisters or task definitively dies), task state would transition appropriately. We might want to do this via a new task state, e.g., TASK_REREGISTER_PENDING.
This might be consistent with changing the task states so that we capture "task is partitioned" as an explicit task state (TASK_UNKNOWN or TASK_WANDERING) – see