Foreign key referential actions

Right now, referential actions get deferred along with normal
checks and run against the state of the database at that time.
I think this violates SQL92 11.8 General Rules 4-6 and have some
reasoning and proposed ideas towards making it more complient
although I don't actually have an implementation in mind for
the most correct version. :(

Here are my interpretations:

GR 4 says that the matching rows (unique and non-unique)
are determined immediately before the execution of an SQL
statement. We can ignore the fluff about non-unique matching
rows for now because I believe that applies to match partial only.
GR 5 says when there's a delete rule and a row of the
referenced table is marked for deletion (if it's not already
marked such) then do something based on the action, for example
mark matching rows for deletion if it is cascade. This seems
to imply the action is supposed to occur immediately, since
AFAICS the rows aren't marked for deletion on the commit but
rather on the delete itself.
GR 6 seems to be pretty much the same for update.

I think the correct course of action would be if I'm right:
*Make referential actions (other than no action) not deferrable
and thus initially immediate. This means that you see the
cascaded (or nulled or defaulted) results immediately, but
I think that satisfies GRs 5 and 6. It also makes the
problems of what we can see a little less problematic, but
doesn't quite cure them.
*To fix the visibility issues I think we'd need to be able to
see what rows matched immediately before the statement and
then reference those rows later, even if the values that we're
keying on have changed. I'm really not sure how we'd do
this without a great deal of extra work.
An intermediate step towards complience would probably
be making sure the row existed before this statement
(I think for the fk constraints this means if it was
created by another statement or a command before this
one) which is wrong if a row that matched before this
statement was modified by this statement to a new value
that we won't match. Most of these cases would be errors
by sql anyway (I think these'd probably be real triggered
data change violations) and would be wrong by our current
implementation as well.

I'm not sure that the intermediate step on the second is
actually worthwhile over just waiting and trying to do it
right, but if I'm right in what it takes, it's reasonably
minimal.