I guess that I have been indoctrinated enough into the concept that side-effects in functions are a bad thing that I strongly avoid writing a function that causes side-effects. (Particularly if it only does it in one context.)

Side effects are okay, as long as they're documented clearly. The efficiency that you get by mutating values instead without copying them first is worth it, in my opinion. I usually document the functions exactly as in my post: a piece of code that calls the function/procedure in all three contexts, each with a useful comment.

As for return @array; - (...) So I left it with the easiest behaviour to implement, knowing full well that it wouldn't do anything useful in scalar context, (...).

Of course, there isn't always a good scalar value. Hence the "unless the function warns or dies in scalar context." in my post. I tend to carp +(caller 0)[3] . ' used in non-list context' unless wantarray;.