Losing IPv6 connectivity after 20min, why?

[Olaf Schreck <chakl@syscall.de>]

I have configured a Debian 7 server for IPv6 (in addition to IPv4). 
I can ping6 www.google.com and other addresses, fine.  BUT the server 
reproducibly looses IPv6 connectivity after roughly 20min, and I can't 
figure why this happens.  Clues anyone?

My hoster (Hetzner) routes the 2a01:4f8:191:XXXX:/64 network to the 
server.  Following their instructions, I assign a static address from that 
block and set the default route to fe80::1, either manually 
 ip -6 addr add 2a01:4f8:191:XXXX::5/64 dev eth0
 ip -6 route add default via fe80::1 dev eth0

or in /etc/network/interfaces like this
 iface eth0 inet6 static
  address 2a01:4f8:191:XXXX::5
  netmask 64
  gateway fe80::1

This works, I can ping6.  But it reproducible stops working after 20min, 
confirmed using this command
 while true; do date; ping6 -c3 -w5 www.google.com; sleep 10; done

I'm sure it's not Google rate-limiting my pings, I get the same results with 
various IPv6 addresses that I'm authorized to ping.

To restore IPv4 connectivity (IPv4 still working), I can either reboot or 
re-add the default route with these commands (order is important):
 ip -6 route del default
 ip -6 route del fe80::1 dev eth0
 ip -6 route add fe80::1 dev eth0
 ip -6 route add default via fe80::1 dev eth0

Which will give another 20min of IPv6. 100% reproducible. And it's just the 
routing that needs to be fixed.

I have exluded:
- no router advertisements used by the hoster, no such packets seen with 
  tcpdump, server is not configured to accept RAs
- no ip6tables rules, and default ACCEPT everywhere
- no cronjob or other periodical script that could be responsible
- no "security software" or similar that would interfere

Important data point: This server has 2 ethernet interfaces, so there are 
2 link-local fe80::/64 routes to eth0 and eth1.  I was suspicious that the 
problem might be related, so I disabled IPv6 on the second interface 
completely with with sysctl net.ipv6.conf.eth1.disable_ipv6 = 1.

And that resulted in stable and flawless IPv6 connectivity!

While this workaround is ok for this server, I have another one that shows 
the same symptoms.  But for that server I need IPv6 on the other interfaces, 
so the workaround does not apply.

I'd rather like to learn why this happens, or what config part I may be 
missing. Clues or further debugging hints very welcome. Thanks!

Olaf