Several commenters have suggested that the discrepancy between the FN vote and the polls shows that people are still reluctant to own up to voting for the extreme right (a version of "the Bradley effect," as it were). But no one has explained why the polls overstated Mélenchon's vote. Were left-wing voters perhaps ashamed to say that they were voting for the bland Hollande rather than the fiery Mélenchon? Is it prerferable on the left to be thought a revolutionary rather than a pragmatist?
A different sort of explanation would look at the polls themselves. French pollsters rely on a method of quotas, which requires considerable understanding of the structure of the electorate before sampling begins. Since the parties in France are unstable, especially the minor parties, that knowledge, based on analysis of previous elections, may be incorrect, leading to biased sampling. I don't know enough about the methods used to say why there might have been systematic errors in favor of Mélenchon and against Le Pen. Perhaps the urban concentration of JLM voters as opposed to the periurban concentration of MLP voters has something to do with it, as demographic patterns have been shifting too. In any case, I am skeptical of the "Bradley effect" stories and would suggest that the flaws are in the polls instead. And this raises doubts about the round 2 polling. The polls were actually quite accurate for the top-tier candidates in round 1, but round 2 depends on getting the transfer of votes right, and the poor results for the second- and third-tier candidates indicates that the polls may not have a good handle on this. So I'm not ruling out a significant surprise on May 6.