Since there is a number of people in this thread, let me take this opportunity to try bust a myth for the larger audience possible.
In all traditional architecture routers, from the smaller 851 to the 7200 NPE-G2, the process does see and process each an any packet no matter if you have configured process, route-cache, or CEF switching.
It is not like CEF or route cache are magic accelerators that do things in hardware. The only difference is that they handle the packet at interrupt time instead that process time.
To understand these terms you need to have a little understating of a typical OS architecture, however we can summarize saying that interrupt time is a routine that starts every single time a packet is received, and is supposed to end as fast as possible.
Process time is all the cpu time left when not doing interrupt-time.
In a very relaxed router there will be plenty of time for both levels, things change when router go under stress. You must prioritize still serve all interrupts to avoid drops, but you can't starve process time because that impacts the control-plane, basically what keeps the router together. That's why we can tune the "scheduler" parameter in some case.
Now that we've introduced interrupt and process level, let's go for an historical perspective.
When they invented the router at the beginning of the internet era, processors were slow and had little memory. So the trade-off was between having few features (to be fast) using a relative small cache of most used routes (to save memory and cpu cycles) versus doing things in a more exhaustive manner against the full route table - process switching exactly.
The idea was that complex features were not apt or too difficult to be implemented at interrupt time - or as you can hear sometime, in the fast path. Packet debug, as rick mentioned before, is one of these.
So they put everything not fitting the fast switching model in the process path, and cisco provided (sometime, not even all the time) a list of what was doable in route-cache, and what not.
Then the processors grew in performance, memory available, and customers wanted to have all features in the fast path, because process switching performances were very disappointing.
Circa the same time CEF was introduced, by some some smart guys that deserves a lot of appreciation. They not only made an infinitely better algorithm to scan an entire route table in less cpu cycles, but also provided programming hooks and restructured code so that most features could be done at interrupt time. So the current tendency of having almost everything done at interrupt time.
Well, I've summarized the matter a lot, I hope that helps understanding things better.
