2) What is the expected (or 'acceptable') gap between R and Rfree at this resolution. (Current R = 27%, Rfree = 33%) Does this sound reasonable?

It's not unreasonable, but you can probably narrow the gap if you get the reference model restraints working. (For comparison, I've spent most of the last week playing around with low-resolution data, and two of my test cases are a 3.6A structure with R/R-free = 0.285/0.335, and a 4.0A structure at 0.27/0.34 - both published, and originally refined with Phenix.)

3) It has been suggested to me that I should try adding riding hydrogens during the last round of refinement (to help with geometry). Is this something I should do?
And if so, should I remove them when I deposit the file to the PDB. (Hydrogens at 4.5A are probably going to raise a lot of eyebrows...

It's worth a try, but in my experience this often just makes it worse.

5) I have been using secondary structure restraints during the refinement, which seems to work reasonably well. I also tried refinement using reference model (the structure of the individual components). But refining with reference models seems to result in high RMSD bonds/angles. Is there something I'm missing?

For some reason the reference model restraints often have a weird effect on the automatic weighting. Try optimizing the x-ray/stereochemistry weighting with relatively strict limits for RMS(bonds)/RMS(angles) - on the command line, for instance:

optimize_wxc=True

bonds_rmsd_max=0.01

angles_rmsd_max=1.5

(and in the GUI, it's in the middle of the second tab in the configuration section.)

-Nat