Details on Convergence are to be found in the Perspectives Project, notably the corresponding Usenix article.
An attacker able to "hide" the server from all notaries (i.e. an attacker working over the local network of the server, mostly) is what the authors call the Lserver model. Ultimately, that attack "works" with default notaries, which use "network perspective". Namely, the notaries will see the attacker's certificate as being "a new certificate for the server", which is valid. However, as soon as the attacker ceases the active attack, the old certificate becomes visible again, and this raises alert flags in the notaries: a certificate update is certainly valid, the reappearance of an old certificate is suggestive of foul play at work.
However, nothing prevents notaries from using other validation strategies, including CA-based validation, which has its own issues but resists MitM.
There are two main ideas in that system:
Delegate validation to external servers (the notaries), using several with a consensus quorum to avoid single points of trust. This kind of delegation already exists in the context of the X.509 bestiary, under the name SCVP (but the Convergence protocol is quite simpler).
Have notaries use their memory to improve security. Where a normal SCVP server would rely on context-free validation (the true X.509 way) with maybe a bit of short-time caching but only to save on CPU costs, Convergence notaries primarily rely on temporal stability: they assume that SSL server certificates usually remain unchanged for one or several years, for a given server; hence they can simply check that a SSL server certificate is still the same than before. X.509 and SCVP could deal with a server getting a new certificate every 5 minutes, which is awesome but completely useless in practice.
A cornerstone of the Convergence reasoning is that if an attacker can successfully defeat the checks performed by the notaries, then he will also be able to impersonate the server when talking to traditional CA, and thus could also defeat the traditional CA model. Hence, security-wise, Convergence is no worse than the present situation. Note that X.509 could do much better; in particular, revocation is not really supported by Convergence, and Convergence is no worse than traditional CA mostly because existing clients tend not to check revocation anyway.
So really Convergence is not about making SSL more secure, but rather about killing off existing CA without lowering the overall security.
Potentially, a notary could run any validation process it wishes, including normal X.509 path validation (and then avoid the kind of MitM that you suggest). Since one could expect that there will be much fewer notaries than clients, and those notaries would be more "security aware", we can conclude that at some future point Convergence would actually improve Web security. The most immediate consequence of widespread Convergence usage, though, would be a shift in power: the Game Master would no longer be Microsoft (who decides which CA go into the list recognized by default by Internet Explorer) but rather whoever runs the notary with the best network connectivity that clients would use by default (presumably Google).