An update.
This issue seemed to be related to the nRF module itself being assigned an incorrect Node address.
After much head scratching I thought it might be something to do with timing, between the nRF relinquishing SPI and the Ethernet shield accessing. SS voltage drop/rise conflicts so that the SPI was still active for a cycle or two when the Ethernet started transferring data. I was guessing that the node settings were being corrupted somehow.
Anyway, I tidied up my demo sketch and inserted a delay between going from nRF to Ethernet and back again. This seemed to cure the problem when I set to 50ms and I gradually worked my way back, splitting 50% each time, through 25ms and 12ms.
Finally I tried no delay and it stills eems to be OK at the moment, so it could have been something dirty in my code after all.
Anyway, it looks like I have a handle on it for now, so thought I'd update here in case anyone else has the same problem and tries to troubleshoot.