Answering my own question..
Not willing to pay "premium price" for boutique cooling products, I ordered 1 mm thick no-name silicon thermal pads.
At first I tried one layer. No change, cache still disabled. I then put two layers (2x 1 mm) of thermal pad and cache briefly came back to life. PowerLogix's CPU Director was able to see it for a moment, but then the machine froze and afterwards the cache remained disabled.
2 mm is definitely too much because with two layers keyboard has a bump in the area where the cache chips are located.
Now comes the interesting part. I was given another machine that had many parts, especially top cover, missing. Powered it up and it had no problems with cache and it did not complain about unknown USB device!!!
Ran it for a couple of hours without the cover - cache chips were only getting warm, not hot, like on the machine with problems.
I also noticed, that the guy with PB G4 5,1 prototype on YouTube had problems with cache.
Protools5LEGuy reported that tightening the keyboard screws - i.e. applying more pressure to cache chips, brought cache back to life.
All this and my experience (I'm an EE) tells me that this is more of a contact than thermal problem. Bad batch of mobos, bad soldering, imperfect soldering jig, failing cache chips or something inbetween. The problem might be similar to failing GPUs on G3 portables or 2011 15" MBPs.
It also might explain why one does not see thermal pads in various videos and don't find thermal pads in machines that have been "serviced" *. Absence of thermal pads will guarantee that the machine will disable failing cache upon startup and nobody, except technically minded people, will notice this in OSX.
* The first one that I bought have traces of "repair" - one of the caps near the cache has been fiddled with.
More later..