DNS responses incomplete when TCP query contains EDNS record

[rade]

$ weave launch
$ weave run --name=foo -ti ubuntu
e14c663b3e5996ac94288a947d5985ff89450ffcdc35e04fde23ca25f2f27d8f
$ for i in $(seq 50); do weave dns-add 10.2.1.$i foo -h bar.weave.local ; done
$ dig +short @$(weave docker-bridge-ip) bar.weave.local A | wc -l
50
$ dig +tcp +short @$(weave docker-bridge-ip) bar.weave.local A | wc -l
50
$ dig +bufsize=700 +short @$(weave docker-bridge-ip) bar.weave.local A | wc -l
22
$ dig +tcp +bufsize=700 +short @$(weave docker-bridge-ip) bar.weave.local A | wc -l
21

Looking at this a bit more deeply, dig by default adds an EDNS OPT RR record indicating the max buffer size is 4k. The +bufsize option shrinks that. dig also falls back to tcp when getting a truncated response.

The reason we get truncation even then is the nameserver takes into account the EDNS-supplied UDP buffer size even when interacting over TCP.

I am pretty sure this is wrong.