I think this is a really great question.
Remember that Kyma X Revealed was written originally for the Capybara and the paca(rana)s are way more powerful. The optimisations in the back of Kyma X Revealed work but with the power of paca(rana)'s I don't find they're addressing my usual causes of Out Of Realtime.
It would be great to collect everyone's tips on optimisation techniques - if you have any do add them to this thread.
Here's a few.
Feedback loops are expensive in DSP resources (using FeedbackLoopInput and FeedbackLoopOutput). You want to minimise the number of loops if you can. A good tip, thanks to Cristian Vogel, is that the FeedbackLoopInput/Output pair is actually stereo when the delay time is set to 1 samp. So you don't need to create two separate feedback loops for the left and right channel. Also you may be able to use a single loop for two purposes using ChannelJoin and Channeller Sounds.
Kyma X briefly mentions manually forcing the processor assignment. It is a fragile technique so one to be wary of but it does work in some situations. Particularly if you've got some Sound that works fine but then runs out of realtime when you try and make copies of it. Say you've got four CrossFilters on a pacarana. If you're just mixing them together at the end then I've had good success using ForcedProcessorAssignment to manually assign each leg to a processor (often using ?VoiceNumber if I'm using a Replicator). But beware if you then try and put anything in the flow after the mix. You may well create a flow that can't be scheduled and you'll run out of realtime again (as the book points out).
Generally speaking you have to try pretty hard to get Capytalk to cause an OORT! There are ways for the determined. TriggeredSoundToGlobalControllers can be handy, but if you replicate a lot of them and they're all gated you can run in to OORT issues. Try changing your Sound so you can use un-gated TSTGCs instead.