In normative text, I don't see how "must" could occur anywhere except
where it was supposed to mean "MUST".
Perhaps, but often we don't clearly distinguish between normative
and non-normative text. Using the 2119 keywords (spelled in all
capitals) helps us do that.
I've certainly seen cases in standards documents where it was useful
to distinguish "should" from "SHOULD". "SHOULD" imposes a (conditional)
requirement on implementations, while "should" can be used for advisory
text.
I don't believe the intent of 2119 was to change the meanings of
"should", "must", etc., in RFCs, but rather to define new terms
"SHOULD", "MUST", etc., with specific new meanings.
Keith