Hi Julien,
this is a normal behavior according the load balancing infrastructure.
Load balancing module will dispatch the job to an available process cluster hence the workflows should be designed to don't stick to particular paths or folders since the process cluster can be any random one according the selection by load balancer.
When you use watch folders they are the source of the initial job but obviously that folder has no meaning on a different physical machine working in clustering, neither the process cluster will ever access directly a source folder being a passive machine receiving jobs by the manager nodes.