IPv6 address still in duplicate address detection is skipped and never retried #1915

Problem

A link-local IPv6 address that is still in duplicate address detection (DAD, "tentative") cannot be bound. new_socket treats EADDRNOTAVAIL as benign, logs a warning and returns None, so the interface gets no responder socket. add_interface then drops the multicast group membership it had just joined for that interface.

Nothing retries the bind. Once DAD finishes (about a second later) the address is usable, but the address set has not changed, so a caller that reconciles on address changes has no reason to call async_update_interfaces again. The instance stays without IPv6 on that interface until some unrelated address change.

Where

• new_socket returns None on EADDRNOTAVAIL.
• add_interface drops the group membership when the responder socket could not be created.

This affects both construction (create_sockets) and async_update_interfaces.

Reproduce

Start a container on an IPv6-enabled Docker network and construct AsyncZeroconf(ip_version=IPVersion.All) as soon as the process starts:

docker network create --ipv6 --subnet fd00:dead:beef::/64 zc6test
docker run --rm --network zc6test <image> python -c "..."

Log output:

WARNING zeroconf: Address not available when binding to ('fe80::943f:67ff:fe01:cbb8', 5353, 0, 11), it is expected to happen on some systems

The instance has an IPv4 sender only. Calling async_update_interfaces() a few seconds later binds the link-local sender.

Seen with 0.151.2 on Linux (arm64, Python 3.13). The same window exists whenever an interface comes up shortly before construction or a reconcile, for example at container or host boot.

Suggested fix

Treat EADDRNOTAVAIL on an IPv6 address as retryable: remember the interfaces skipped for that reason and schedule one short internal retry (a few seconds), for both construction and async_update_interfaces.

Filtering tentative addresses up front is not an option, since ifaddr does not expose IFA_F_TENTATIVE.

Workaround

esphome/device-builder#2864 runs one extra async_update_interfaces ten seconds after startup and after each binding change.