Thank you for the detailed description. Here is my suggestions:
1: do not reduce dt factor if you have not identified the cause of divergence. Only when using Metal BCS and it is still diverging then you can reduce it.
2: The distance of PML is good, in most cases except there is long-tail SPP or evanascent wave. You can use a profile monitor to check the field.
3: when simulating periodic structure, most likely it has higher order diffraction. Even though there is no higher order diffraction in the wavelength range you are interested (mostly the source bandwidth), it does not prevent large diffraction angles from shorter wavelength in the source, outside of the interested bandwidth. You can use a transmission/reflection monitor to record shorter wavelength, and check the farfield by using Visualizer with about 10 periods (approcimately. or you can use grating analysis). You may need to use log scale to find the higher orders.
I guess it is most likely due to #3. You can
1: set the source in time domain, using longer pulse and offset. For example 50/100fs:
2: using Steep Angle PML with more number of layers