In fact try the following variation of the eval and tell me whether it works better

You the monk!

Yes, it does. And, because of your patient explanation, I understand better what is going on. A few questions niggle my mind and remain --

1. When I convert the warn into a die it gets trapped by eval. Why? I mean, why doesn't die behave the same way as warn... it is, after all, an error, and should be redirected to STDERR. Mind you, the way it is working _is_ how I want it to be. I just don't get this discrimination.
2. When warn gets piped to STDERR, why does the script stop working? After all, it is just a warn and not a die. I would expect the script to warn and then continue working merrily. I've put it in an infinite loop with a sleep statement in there, so it should just take a breather and then try again. Except, if I don't modify eval as you have suggested above, the warn goes to STDERR and the script quits.
3. Since the above modification works, it should also work if instead of warn I explicitly use die. In fact, I am going to try that right away and see if it does so. In that case the local modification of SIG{__WARN__} should not be necessary.
4. And finally, I am assuming I can convert other SIGs the same way. In fact, I don't want the script to die ever unless CTRL-C/BREAK is pressed (SIGINT) or the computer is shutdown. Any caveats to that?