After fitting a PDB into a 3.5-5 A resolution EM map by phenix.real_space_refine, should I still need to do a reciprocal space refine?

Why?

For crystallography, the reciprocal space is in fact not purely imaginary,

Indeed!

But for the EM data refinement, what is the reciprocal space refine?

If you imagine cryo-EM data to be in reciprocal space you can do reciprocal space refinement then. This may work given you know why you do this.

Pavel