'If you can store infinite information outside of a system, then you can achieve infinite information density inside the system if you use outside the system as 'context'. I hope that paraphasing makes it clear enough the problem with this line of thinking. You have to include all the information when calculating absolute density. In your case, the context has to be stored somewhere too.
If the outside system is a probability distribution over output messages (a more general case of the dictionary you described) then the problem is synonymous with compression.
Shannon information is always measured relative to a receiving context in which it the symbols are understood, and information content is related to the inverse of the probability of observing a particular signal as assessed by the receiver. So from that perspective vinceguidry has a reasonable point.
However really the question being asked is something more like 'data density' and that is generally what people are talking about when the term 'information density' is invoked.
Edit: I see that the original article does indeed refer to Data Density, and that HN title is just wrong.