Yes, I suggested that class authors not return $self from mutators. Of course, as the discussion in multiple method calls against the same object, revisited shows, there are ways to retrofit chaining from the user's end to classes that don't implement it (I was almost going to upload one of those to CPAN).
So I don't see a valid case for returning $self as an enabler of choice: the choice is always available, regardless of the classes interface. What's left then is the question of encouraged usage style (see the docs to SOAP::Lite). And to me, there's no question that this style should not be encouraged, just as there's no question that use of goto should not be encouraged, regardless of one or the other is sometimes the right tool or not (for which there is clear evidence in case of goto; with mutator chaining I've yet to see any).
The only reason to favour chainable mutators I've seen anyone mention is brevity, conciseness. But as a class author you can offer that to users without conflating semantics by offering a single unified setter.
Makeshifts last the longest.