Granted, not everything listed stops at unspecified behaviour. I'm not convinced that that's a good thing, though. Even something like giving signed integer overflow unsigned semantics is pretty effectively useless. Sure, you can reasonably rely on twos-complement representation, but that doesn't change the fact that you can't represent the number six billion in a thirty-two bit integer, and it doesn't make 32-bit code that happens to depend on the arithmetic properties of the number six billion correct just because the result of multiplying two billion by three is well-defined.
Then there's portability. Strict aliasing is a good example of this. Sure, you can access an "int" aligned however you like on x86. It'll be slow, but it'll work. On MIPS, though? Well, the compiler could generate code to automate the scatter-gather process of accessing unaligned memory locations. This is C, though. It's supposed to be close to the metal; it's supposed to force-- I mean, let you do everything yourself, without hiding complexity from the programmer. How far should the language semantics be stretched to compensate for programmers' implicit, flawed mental model of the machine, and at what point do we realize that we already have much better tools for that level of abstraction?