Most likely not (but almost). Apart from the specific papers without references that people have already mentioned, we can look at crawls of large collections of papers. Take, for example, this dataset of all papers in the theoretical high-energy physics section of ArXiv:
Nodes in largest WCC 27400 (0.987)
The largest weakly connected component (WCC) is what you're after: the largest subset of papers that are connected to each other by a path of citations (ignoring direction). While the largest WCC is almost as big as the entire graph, there are papers outside it. Usually, with graph like this, these form little clusters of their own.
For a more cross-domain dataset, consider the citeseer graph, again a small proportion of papers, outside the largest WCC.
Now, of course, these datasets don't contain all of academia, and adding more papers would mean connecting some islands to the WCC, but I'd say adding more papers also adds new little islands. No matter what rule you use to decide which papers count and which don't, I think you always end up with disconnected islands.
Of course, if your question is whether any randomly chosen paper in domain A is likely to be connected with a random paper in domain B, the answer is yes. There will be a large WCC encompassing all domains, and a few tiny islands. I've seen visualizations to this effect, but unfortunately I can't find them at the moment.