ciphergoth: The argument you put forth applies to some types of trust metrics, but not the one I gave.
Specifically, I think your argument applies to trust metrics in which first compute a weight for each attester, then combine their judgements by simple averaging. The second step of the algorithm I gave is considerably more complicated. You imagine doing a random walk around the graph, with each hop to one of the ones certified by the current node, weighting the chances of going to each one proportionally to its confidence. The histogram of what this process will result in can be averaged in any way you choose, median is probably best.
If an attacker has compromised some nodes, their most effective attack will be to make all those nodes attest to what they want. Adding lots of nodes and certs can only asymptotically approach that amount of influence.
After thinking about it a bit, I realized that sampling works quite well as a method of computing confidences. Randomized algorithms show up in the most surprising places.