So if you write to a persistently mapped resource in a command list and wait for a signal, you're guaranteed to have the writes finished without any barriers? That seems odd because shouldn't the signal only be informing the cpu of the current execution state?

If you write to a persistently mapped buffer in a command list, and wait for a signal, yes the writes will complete without any barriers on the buffer. For textures in CPU-accessible heap types, the resource needs to be in the COMMON state for the CPU to read it.