Precisely. This optimization is possible because according to ANSI C, a variable is allowed to point to an address anywhere within an object, or at most one (sizeof) past an object. Anything else is undefined behavior (what they refer to as as "illegal pointer" in the post).

Therefore, buffer through buffer+1 would be valid pointers, but unless they can prove that i-128 is <=1 they can safely assume that the final pointer can be more than 1 past the buffer.

If it was just a pointer (not an array) they could do the same analysis if the pointer was mallocced right then in the same function with a constant size. More likely though, the pointer would be a function argument and then they would not know where the object boundaries were without doing interprocedural analysis (which I assume is outside of the scope of this post) and some other additional propagation (so technically within the C99 spec, but I doubt you would see an optimization that actually would do that).