Also, endian-correct code is usually semantically clearer. For example, if you're reading network-ordered bytes into an int, an unconditional endian swap (which will produce correct results on LE systems but not BE) is less clear than invoking a "network bytes to u32" helper.

There are a lot of odd (by modern standards) machines out there.

You're also the most likely person to try to run your code on an 18 bit machine.