To summarize some off-list dicussion - the kernel guys have said that using sendfile on IPv6 sockets may trigger bugs in cards which do hardware TCP checksumming for card/driver/OS combinations which support that. (since the cards have to know about IPv6, and that probably doesn't get tested much).
Colm confirmed using tcpdump traces that the "missing" response packets were indeed getting sent with bad checksums. ("tcpdump -vv -s 1500 ip6" is the magic invocation for the curious) joe