There does not appear to be any incentive to run a full node (or indeed, a (lowly ranked) standby witness).
From the bits and bobs I remember, memory usage on witness/full nodes is still rising linearly, altough there was some sort of patch around to drastically reduce this?
Some of the documentation talks about the benefits of running a full node for decentralization - correct assumption that running an actual (idle) witness for the same purpose accomplishes exactly the same?
And finally, could a witness operator comment on the require hardware to run a witness/full node ?