SOLVED. Short version: my hosting provider (Hetzner) obviously fixed something *silently*. Everything works as initially configured, no workarounds required.
Please disregard any speculations I made in this thread, especially that disabling IPv6 on some interfaces or fiddling with link-local fe80:: routes might be useful. These were conclusions based on coincidence rather than causality. Many thanks to Marc for very useful debugging hints, to Gertiaan for pushing to read NDP specs, and to everyone else who bothered to reply on this dull problem. No thanks to Hetzner, for either lying ("we didn't change anything") or not having a clue. I did a lot of debugging with ping6, tcpdump, rtmon and syslog. It became quite obvious that the problem was at the uplink router which seemed to "forget" the /64 route after 20min. I provided some debugging hints in a support ticket, and suddenly, magically everything worked without further changes on my end. "We didn't change anything". Yeah, sure. Olaf