Could I / we test this? Send some fragments to your favourite server, and arrange to receive some?
I could of course reduce my own MTU, but that might well simply cause things to adapt and reduce packet sizes anyway so I simply get unfragmented smaller packets as PMTUD does it's good thing.
And now I think of it, unfortunately I don't have the tools to inspect the results at the moment either as I'm confined to my main operations base these days.
I would be also very interested in how firewalls deal with this in general. (i) Anyone know anything? But (ii) also especially interested in the case where an initial fragment is suspiciously short, say if an L4 header is split across a fragment boundary, this might look like an attempt to evade security checks by obfuscation? (iii) Do firewalls buffer fragments in case ii then, waiting until they have enough contiguous data to get an entire L4 header to inspect it or however much of the packet they need depending on how deep their inspection is? (iv.) What happens if fragments arrive out of order at a firewall?
Makes you think there ought to be a standard test program for generating a set of test cases to check that firewalls, routers and hosts correctly handle fragments and we should be able to give various classes of boxen pass / fail stickers, or nice little stars and scores out of 100 for correct behaviour. (Following on from the success of the AcidX test for CSS capabilities and compliance in web rendering engines.)