Oh my god, what a dumb dumb I am. Penny has dropped.
Now getting 184pps from a cheap dual-core G2020 ivy bridge, using original unmodified sourcecode. Pps still fluctuates though.

Ahh, can't believe it's take me over 2 days to realise what I had to do to double the pps. Kicking myself.

The primespersec measure is for fun only, it does not necessarily reflect true mining performance. The only accurate measure is block rate, that is, given a fixed difficulty how fast can you generate blocks. To measure this one can use a lower difficulty than the network difficulty.