I think you misunderstood the example -- the memory is cleared after use to ensure that if it's reallocated by someone else, or someone hooks up a debugger, the content can't be examined (except when the compiler removes this clearing attempt because of an optimization). Lets say that chunk of memory held a password -- you'd definitely want to clear it after use, even if you immediately free it and never plan to read it again.

That's actually a very good example, but I'd argue that this is actually a violation of the standard: memset is defined as setting the value in memory. Most optimizations on undefined behavior don't really fall into this category.

I guess you could group this kind of thing into the category of "dead code elimination" which is useful, but results in parts of the code written not producing the specified executable. I have to think on this example more.

It is allowed under the "as if" rule. If no visible aspects of the program are changed by an optimization, then it is allowed. The value stored in memory is not considered to be a visible aspect, and so the compiler is allowed to modify which memory is changed.

It's the same as inlining a function. The standard says that a function call is a function call. Compilers are still allowed to inline the call, even if it has not been specifically marked as "inline".