My understanding has always been that the substring of a string should share its utf8'ness, and that once turned on it stays on until explicitly turned off.

It makes sense to me that once turned on in a particular string, it should stay on (e.g., if the string is modified via $str =~ s///). Functions that give you substrings (e.g., substr, split) create new strings for these, so there is no "staying on" to be done. The flag value would have to be intentionally propagated from one string to another.

Even a basic assignment creates a new string, but one would hope one of the properties of an assignment operator is that it duplicates both a variable's data and its metadata. (Yet even this fairly straightforward fact is not documented in perlop.)

Anything else I consider a bug.

By that logic, the behavior hv points out in Re^7: Seeking Perl docs about how UTF8 flag propagates is a bug. But since no documentation supports your expectation, I'm not sure you could make a case for that.

Having said that, it is good form to treat it as an uncertain value

Yeah, that's what I'm doing now, in response to this thread. Thanks to everyone who's chimed in.