You've mentioned NaN propagation in another comment and I think that's a perfect example of the problem with this approach. Sorting a vector of arbitrary floats is a notoriously thorny problem because any float could be NaN, and as NaN is incomparable to any other float, there is no total ordering of floats. There is no general solution to this problem that doesn't involve making assumptions that could be faulty for some applications.

There are sensible answers. But they are weird.

I would say let an INVALID string be length 0. Then accept that catting a valid and invalid string would result in a shorter length.

Which one do you think is safer?

Negative lengths are not compatible with unsigned representation.

A system implementing invalid string values must choose a text encoding such as UTF-8 that supports the concept of an invalid character. Null termination is too flexible. As such is simple length prepending.

I understand how one could shoot down implementations, but none has made a convincing argument about shooting down the idea.

memcpy and company are strictly for raw unencoded buffers.

C doesn’t have the notion of “size of buffer” (yes, arrays have a size that can be queried by sizeof, but only at compile time). You would have to fix that, first.

Isn’t that just NULL?

Let's take strstr, which finds a matching substring needle in a haystack string.

-returns a NULL string if the needle is not in the haystack. -returns pointer to first matching substring.

Extend strstr with VALIDITY

Understood behaviour if both are valid.

Say the haystack is INVALID...as the return value is NULL or a strict substring of haystack, should return INVALID. A poison haystack should poison dependent strings.

Say the haystack is valid but the needle is INVALID...should return NULL. A valid string never contains an INVALID string as a subsequence.

  strstr(NULL, /* valid string */)

  strstr(/* valid string */, NULL)