If you did some tests with ML ceramic X7R caps as are often specified with switchers we would then have real results we can use.
That was exactly my thought. Electrolytic caps have essentially no ability to filter high frequency noise and transients (for some definition of "high"), so there was no way the circuit was ever going to be quiet with just an electrolytic output cap. Stretching out the circuit over quite so much board area isn't helping either, even though the dead-bug style of construction is otherwise pretty good.
I've used the 7660 and generally found it does a good job, but I'd always use it in surface mount form with X7R (or maybe X5R) ceramic caps located right next to the device.
I'd suggest starting again, but instead of leaded caps, solder some SMT caps directly between the pins of the 7660 and the ground plane. No leads == minimal series inductance. Something like this should work well and shouldn't be too fiddly to work with:
http://uk.farnell.com/murata/grm31cr71e106ka12l/cap-mlcc-x7r-10uf-25v-1206/dp/1828837I'd expect the result to be not too far from the multi-stage filtered method in terms of performance, and about 1/4 the size.