The ban logic is per source IP, so it only works where the daemon can see
the real client. Behind Docker's default bridge networking every client is
SNAT'd to the bridge gateway (a 172.16/12 address), so a single IP would
stand in for the whole internet -- counting offenses against it would block
everyone at once.
Add is_bannable_address(): only globally-routable unicast addresses are
tracked. Loopback, RFC1918 private, CGNAT (100.64/10), link-local, IPv6
unique-local, and multicast all return false. main.cpp decides trackability
from the accepted endpoint and skips both the block check and offense
recording for non-global sources. Net effect: banning works where the real
IP is visible (FreeBSD jail via pf rdr; Docker with host networking) and is
inert -- not catastrophic -- where it is not (Docker bridge).
Document the Docker client-IP caveat: docker-compose.yml now defaults to
host networking, with the rationale and alternatives in DOCKER.md.
The Alpine runtime stage uses BusyBox nc, which has no -z flag, and
'localhost' resolves to ::1 while finger only binds 0.0.0.0:79. Both
combined caused every healthcheck to fail (35k+ failing streak on
mammut). Use '-w 1 127.0.0.1 79 < /dev/null' instead.
pmb