Now that I have multiple examples, I have better support for why mechanical analysis of a question body, via string token checks, is horrible at trying to divine the intent of the question author. The only element that works rather successfully is the human element.
Which we already fail at concerning performance, because we turn up a lot of false positives using that as well! Speaking from personal experience, I got a lecture from Eric Lippert on premature optimization when all I cared about was actual significant mechanical issues like program breakdown or leakage or anything.
Mechanical checks would be worse because it is subject to cyclic degeneration. First you will catch all the people who state "I am not asking about optimization" as a false positive. Then you'll miss those users who miraculously have the ability to ask these questions without ever mentioning those words. The word catcher is updated to combat these problem users, which in turn produces even more false positives. You'll even start to catch users who aren't remotely asking about performance. Update the algorithm enough and it becomes impossible to ask a performance question without getting hit, but legitimate users will be inundated with warnings that are simply not applicable.
I imagine that it is very uncomfortable to have to deal with all of the bad performance questions that show up. But I don't think the maxim of "misery loves company" is really the best solution here. Not to mention, having someone walk you through why premature optimization is bad, or explain using a profiler, would be much nicer than having a machine plaster it in your face.