And something I wanted to mention before but forgot to do - there is not only a problem if the responsible writer fails while trying to rotate the buffers but also if there is just another writer trying to write before the the buffer rotation completed. There is really not much you can do in this case besides retrying until you succeed. But this again also means that a writer may have bad luck and every time he looks a buffer rotation - not necessarily the same one - is in progress causing the writer to starve.
And just in case I wasn't clear about that, I am looking at all of this from a pretty theoretical point of view - I have at best a very rough idea what this will be used for in the real world and what scenarios are important and likely and what just doesn't matter. Actually maybe even the thing I just wrote is nonsense because nobody in your target audience will simply kill threads, but then again developers are notoriously good at bending and ignoring rules.
Each thread under the algorithm can perform their actions in a finite number of steps without ever blocking. This means the producers can continue to do other work. The consumer can continue to consume from other log buffers without being blocked and complete in a finite number of steps. If a producer is killed mid operation then no further progress can be made on that log buffer. If this is considered blocking then the algorithm is blocking and therefore not wait-free. It would need to be killed by another malicious thread for this to happen.
What is clear is that this algorithm gives the best latency profile of of all the measured messaging systems and the highest throughput. I now have the challenge of searching for a name that best describes its behaviour.