You don't even need that length prefix if you use something like SDNV. Ironically this would have introduced a case where you have something like a null terminator that can overflow.

It would have also been a fair bit of overhead on those old single-Mhz machines. Still, almost everything you do with strings is slow so it probably would have been manageable. As an added bonus, it is zero space overhead for short strings (up to 127 characters) which are the majority of strings.

In the PDP era, they would have noticed the overhead of having a variable-length int at the beginning they would have to decode.

In the modern era, it's probably cheap to the point of being free, because in the vast majority of cases I would expect branch prediction to largely eliminate the checks as being very predictable.

Even in the PDP era I am not so sure, given the architectures and OSes being developed outside AT&T, some of them in production.

So then....they chose correctly?

Despite being vigorously opposed to NUL-termination in the modern era, yes, I would not criticize the people who actually made the decision for the PDP. They had no real reason to believe we'd still be discussing that decision 60 years later.