Since the discovery of the problem, I found out that it somehow was connected with the machine going to sleep, because it seemed that after waking up from sleep the problems occurred, but not after a fresh reboot.
I moved my OpenCL endeavors to a another, Linux based, machine since then.
I'm now on 10.8.2, XCode 4.5 (updated from 10.8.1, fresh install of Xcode from developer.apple.com not via AppStore) and just ran some tests with succes, even when waking up from sleep. I couldn't reproduce now.
Unfortunately I don't know what exactly helped the problem for the time being. My OSX environment is complicating the issue, as I have one admin account to install applications, and a development account that only has default user permissions. Because of this I ran into compilation problems with other non-OpenCL projects (forcing me to run `sudo xcodebuild -license` on development account, after `su admin`, without sudo I had some problems). What I see when I first started XCode 4.5 (fresh install!) from the development account was a dialog asking for admin permissions to install "Device Components" (whatever that is). So maybe some of that helped.
I observe a similar issue, which I reported to Apple a while ago (radar #1218677. Maybe this is related to your problem:
Simply creating a context and a command queue, and then building a kernel and finally releasing everything again works on a new MacBook Pro Retina w/ Nvidia 650m only ~25 times. After that I get an OpenCL error -30 and I have to reboot the machine in order to be able to create a command queue again (in any OpenCL app). This can be reproduced on several rMBPs on 10.8.x, but not on other Macs.
I have a simple test tool, if you are interested.
FYI Apple confirmed they could reproduce the problem. NVIDIA so far did not yet deliver a fix, though.
Resurrecting this thread, please post back here if you guys find a solution to the problem. This is affecting a product that I am building for the Mac App Store