See your sibling comment for another perspective! (EDIT: which, to be clear, doesn't invalidate your point. Though it's worth considering, I guess, whether "only assigned semantics by the host specified in the domain" prevents user-tracking systems from calling "foo+bar@gmail.com" the same user as "foo@gmail.com". After all - if they're being interpreted "as" user IDs, rather than as emails, does that really breach the RFC?)
It's not really a different perspective. Sieve, which the sibling comment's RFC extends, is a mail-filtering script language for end-user inboxes. So it's perfectly reasonable for a user on foo.com, who knows that foo.com supports the `+` syntax, to write a Sieve script directing mail to "username+blah@foo.com" to a particular inbox.
In fact, that RFC specifically calls out that interpreting the `+` on non-local addresses is likely wrong:
> NOTE: Because the encoding of detailed addresses are site and/or
implementation specific, using the subaddress extension on foreign
addresses (such as the envelope "from" address or originator
header fields) may lead to inconsistent or incorrect results.
EDIT to address your second point:
> After all - if they're being interpreted "as" user IDs, rather than as emails, does that really breach the RFC?
Well, technically no, the RFC is about SMTP so if you're not writing an SMTP implementation, you're not breaching it.
But RFCs aren't the law, so whether you're technically breaching it isn't really what's relevant. What _is_ relevant is that a system that treats foo+bar@quux.com the same as foo@quux.com is making assumptions about how email works that contradict the RFCs that define how email works. Whether that's a useful thing to do in practice is an engineering decision with tradeoffs. E.g., it's probably fine to assume it for a whitelisted set of domains where you know it to be true, like gmail.
> the local-part MUST be interpreted and assigned semantics only by the host specified in the domain part of the address.