> May I toss in a possible complication?   I'm thinking of 
> numeric code, and
> the possibilities of optimisation.  

I'm not saying that you can't get better performance
if you leave things unspecified.  Rather, a programmer
might like to know whether they'll get 0 or NaN (or
an exception) rather than "I'll get one of those things
real fast" :-)

Perhaps a practical solution would be to have a flag
controlling whether or not evaluation order is fixed.
Speed demons can turn the flag off, teachers can leave
the flag on.  

It wouldn't be that hard to implement, either -- leave
the compiler largely alone, but do an A-normalization
pass introducing explicit let-expressions to force
the evaluation order.  

-Greg