As shown in How the Distributor works, a worker checks in for work with the distributor by sending a register message to it. From that moment the worker starts receiving messages from the distributor.
If a worker goes down, messages can get stuck in the distributor or the worker.
The distributor forwards the message to the worker and relies on the store and forward messaging provided by MSMQ to make sure that it gets there. The message is stored in an outgoing queue and processed when the worker returns to working order.
The solution to a "broken worker" or a "worker node's hard disk going up in smoke" is virtualization, where worker nodes run in a virtual machine whose image is stored redundantly in a SAN. That way any messages stored locally in the worker node end up physically stored in the image on the SAN. So if the host machine fails, the hypervisor can bring up the virtual machine that went down on a different host machine, and all the messages can be processed.