> I expected the given example to produce assembly that respects short-circuiting, [...].
> Always assumed this was a requirement of the language, not "up to the compiler".
But the assembly does respect short-circuiting in C++!
The language specification tells you how to figure out the meaning of a program and it tells you which behaviors you are guaranteed to observe or not observe from the program [0]. This is completely abstract – we could ourselves carry out a C++ program in our heads, or with paper and pencil.
A compiler takes a program as input and writes code for some particular machine as output. If the guaranteed behaviors of the program occur when the machine runs the code, the compiler has conformed to the language specification [1]. The language specification says almost nothing about what this code has to look like.
Even if the language specification states that `b == 0` is not evaluated in `a == 0 || b == 0` when `a` is zero (and that this is required, regardless of the right-hand side), the corresponding assembly might always involve executing a `cmp b, 0` instruction, because what it means to "evaluate b == 0" has to do with the meaning of the C++ program, not the assembly.
[0] See the definition of "observable behavior" quoted in https://stackoverflow.com/a/15718279.
[1] See https://en.cppreference.com/w/cpp/language/as_if. This actually detracts somewhat from my point: talking about allowable transformations implies that there's some preferred way of initially compiling a program, some basic version that's then optimized. Even if compilers happen to work this way, the language specification doesn't.