Signed integer overflow is the most common one, since it means you can hoist 32bit ints into 64bit ints so they fit in one register or similar and save hitting the stack.
I was about to say that "this would break a ton of other projects"—but actually, it seems like it'd be fine as long as it was an opt-in -W switch.
Let's say that, on the pathological compiler, it's documented that sizeof(int) is 4, and chars are signed if the number of seconds in the current minute is even, otherwise sizeof(int) is 5 and chars are unsigned. If your code compiles and works with either setup, it's probably correct and portable.
It's not really feasible for programmers to avoid implementation-defined things like sizeof(int) varying across implementations, not is it feasible for compilers to detect assumptions made about the implementation's definition in the program. The existence of a pathological compiler would make it far easier to write programs that work across all implementations.
It's true that working with unpredictable systems is a nightmare, but that's precisely the nature of undefined and implementation-defined behavior in C. The existence of a pathological compiler would merely allow developers to work through these issues on a single machine, which would be an improvement on needing to obtain a variety of compilers and platforms test with to find these issues.