I haven't thought about this too hard, but why doesn't a good bomb act like an observation, and collapse the wavefuntion prior to points C or D?
It does...and that's precisely why this scheme works!
In the "dud" case, there's no measurement, so there's a superposition of paths: the single photon takes both the clockwise (or upper) path AND the counter-clockwise (or lower) path. The two superposed wavefunctions cancel at C and constructively interfere at D.
In the "good bomb" case, there is a measurement, as you point out: either the bomb goes off, telling you it took the counter-clockwise (lower) path, or it doesn't, telling you it took the clockwise (upper) path. Those two possibilities are not superposed, so they can't interfere. The photon is forced by the measurement to take either one path or the other, but not both. But if it takes the pure clockwise (upper) path, it still has a 50% chance of being reflected to C by the half-silvered mirror in the C-D corner. So in 25% of the "good bomb" cases, you see no explosion, but measure the photon at detector C, which can never happen in the "dud" case. (50% of the time, the bomb goes off, and 25% of the time the photon ends up at detector D, which looks just like 100% of the "dud" cases).
Actually, all the bombs explode - that is, in other worlds. According to the MWI, you kind of cheat the universe splitting, and you get to keep some good bombs that DID explode in another universe.