For a nanofactory to work, a lot of different things have to come together. So why are we so confident that it will work?
Most of the things that will be necessary have been thought about enough to know that they won't be showstoppers. But first, let's acknowledge what's not yet certain:
A complete set of diamond-building reactions has not yet been developed. On the other hand, there are a lot of different ways to make diamond: under high pressure, from low pressure gas, and by irradiating bucky-onions. The current research goal is to add carbon to un-terminated surfaces, but this is making it hard on ourselves. Putting hydrogen on the surface, and removing it just where new atoms need to be added, should keep the surface much more stable.
And, in the unlikely event that there's just no way to build diamond mechanically, there are other materials that are almost as high performance. The chance that all of them will turn out to be impossible is very small, and certainly that uncertainy is no excuse for failing to prepare for molecular manufacturing--especially since we know that biology is possible, and offers a lower bound for nanomachine and nanofabrication performance.
By contrast, there are a lot of implementation aspects of MM that haven't been built yet, but we don't really need to worry much about whether they are feasible. For example, the software to pass data through a nanofactory need not be extremely complex. (The CAD software to help humans enter product designs is another level of difficulty, but still something we can do.) Even a primitive and inefficient nanofactory can be cooled.
There are several different classes of problem. Some kinds of problem form the physical context in which everything will have to be designed. Large machines have to cope with gravity. Nanoscale machines have to cope with thermal noise (and to a much lesser extent, quantum uncertainty). A number of molecular machines have been simulated mechanically, and it seems pretty clear that nanoscale machinery can be designed to work under nanoscale conditions.
Some kinds of problem are just engineering. For example, how will the nanofactory be laid out physically so that subcomponents can be transferred efficiently from place to place, while feedstock, power, and cooling fluid are distributed to large numbers of fabricators? Several nanofactory designs have been developed, and it appears that this particular problem need not even be very difficult. The design of all the robotic components inside the nanofactory--the interior decorating job, if you will--will be a rather large engineering task. But it does not appear more difficult than building a comparable set of macro-scale robots, and there are some factors that may make it easier--like lack of gravity and wear.
Some problems are really non-problems. A lot of nanotech deals with the interesting behavior of electrons. But molecular manufacturing is almost entirely mechanical, and the electrons stay attached to their atoms and behave in a boring and predictable fashion. Likewise, a lot of chemistry deals with very small and subtle differences in energy, but molecular manufacturing can handle much larger energy differences, easier to understand, simulate, and design for. Intuitions from other nanoscale fields will often make the problems seem much harder than they are.
So... when will we know for sure that a nanofactory is possible? To a determined skeptic, the answer is: when one is built, and not before. To an open-minded skeptic, a reasonable position may be: when a full set of diamond-building reactions is worked out. But having talked to a lot of skeptics, it seems that their skepticism decreases the more they learn about how molecular manufacturing works in detail. Also, it should not be forgotten that skepticism has been exacerbated by demonstrably incorrect claims made by some scientists. In other words, the prevailing level of skepticism is higher than it would be if skeptics had more information and less misinformation, and the typical estimate of difficulty or unlikelihood is probably an overestimate.