fix: surface avahi .deb install failures, and skip eth0 smoke check without a cable

.github/workflows/validate.yml

@@ -32,11 +32,22 @@ jobs:
               install -d -m 700 /etc/NetworkManager/system-connections
               install -m 600 /src/files/dpea-eth0.nmconnection \
                 /etc/NetworkManager/system-connections/dpea-eth0.nmconnection
+              # Simulate the real RPi OS Desktop base, which already ships stock avahi.
+              # A bare Trixie container does NOT, so without this the avahi-dpea
+              # file-overwrite conflict (a missing Replaces:) stays hidden here and only
+              # surfaces on the real image build. Installing stock avahi first makes CI
+              # reproduce that conflict, which is the whole point of this gate.
+              apt-get install -y avahi-daemon libnss-mdns
               bash /src/customize.sh
               echo "===== assertions ====="
               fail=0
               command -v uv >/dev/null && echo "OK   uv $(uv --version)" || { echo "FAIL uv"; fail=1; }
               dpkg -s avahi-dpea >/dev/null 2>&1 && echo "OK   avahi-dpea installed" || { echo "FAIL avahi-dpea"; fail=1; }
+              # It must actually take OVER the stock daemon binary, not just install
+              # alongside it. Replaces: transfers file ownership to avahi-dpea.
+              dpkg -S /usr/sbin/avahi-daemon 2>/dev/null | grep -q '^avahi-dpea:' \
+                && echo "OK   /usr/sbin/avahi-daemon owned by avahi-dpea" \
+                || { echo "FAIL avahi-daemon binary not replaced by avahi-dpea"; fail=1; }
               for line in "dtparam=i2c_arm=on" "dtparam=spi=on" "enable_uart=1" "dtoverlay=disable-bt"; do
                 grep -qx "$line" /boot/firmware/config.txt && echo "OK   config: $line" || { echo "FAIL config: $line"; fail=1; }
               done

customize.sh

@@ -1,9 +1,12 @@
-#!/bin/bash -e
+#!/bin/bash
 # DPEA image customizations. Runs INSIDE the mounted image chroot (see build-image.sh),
 # and inside the CI validation container (see .github/workflows/validate.yml).
 # Expects /tmp/packages.txt already copied in, and the NetworkManager profile
 # already placed at /etc/NetworkManager/system-connections/dpea-eth0.nmconnection.
 
+# Both callers invoke this as `bash customize.sh`, which ignores the shebang's
+# flags, so set the strict flags in the body or a failed step is silently skipped.
+set -euo pipefail
 export DEBIAN_FRONTEND=noninteractive
 
 # --- shared non-Python system packages ---
@@ -25,7 +28,17 @@ grep -q '^i2c-dev' /etc/modules || echo 'i2c-dev' >> /etc/modules
 AVAHI_DEB_URL="${AVAHI_DEB_URL:-https://github.com/dpengineering/avahi_0.8/releases/latest/download/avahi-dpea_0.8_arm64.deb}"
 if curl -fLsS "$AVAHI_DEB_URL" -o /tmp/avahi.deb; then
 	apt-get install -y /tmp/avahi.deb
-	echo 'avahi-daemon hold' | dpkg --set-selections || true
+	# The stock RPi OS Desktop base already ships avahi, so the .deb must overwrite
+	# its files (it declares Replaces: for them). A dpkg file-overwrite conflict can
+	# leave the package uninstalled while apt still exits 0, so assert it is actually
+	# present instead of trusting the exit code.
+	dpkg -s avahi-dpea >/dev/null 2>&1 \
+		|| { echo "ERROR: avahi-dpea did not install (dpkg file conflict / missing Replaces:?)" >&2; exit 1; }
+	# Refresh the linker cache so the patched libavahi-core (mDNS on 5358) it just
+	# dropped in is the one the daemon loads, not the stock lib it overwrote.
+	ldconfig
+	# Hold both so an apt upgrade cannot restore the stock 5353 build over ours.
+	apt-mark hold avahi-daemon avahi-dpea >/dev/null 2>&1 || true
 	rm -f /tmp/avahi.deb
 else
 	echo "ERROR: prebuilt 5358 avahi .deb not found at $AVAHI_DEB_URL" >&2

hardware-smoke-test.sh

@@ -5,19 +5,27 @@
 # real device nodes, the serial mapping, live mDNS port, and the eth0 address.
 
 set -u
-pass=0; fail=0
+pass=0; fail=0; skipped=0
 chk(){ if eval "$2" >/dev/null 2>&1; then echo "PASS  $1"; pass=$((pass+1)); else echo "FAIL  $1"; fail=$((fail+1)); fi; }
+skip(){ echo "SKIP  $1 ($2)"; skipped=$((skipped+1)); }
 
 chk "uv installed"                "command -v uv"
 chk "I2C device node present"     "ls /dev/i2c-*"
 chk "SPI device node present"     "ls /dev/spidev*"
 chk "/dev/serial0 -> ttyAMA0"     "[ \"\$(readlink -f /dev/serial0)\" = /dev/ttyAMA0 ]"
 chk "avahi listening on 5358"     "ss -lun | grep -q ':5358'"
 chk "avahi NOT on 5353"           "! ss -lun | grep -q ':5353'"
-chk "eth0 holds 172.17.21.2"      "ip -4 addr show eth0 | grep -q '172.17.21.2'"
+# The eth0 static (172.17.21.2) only activates when the link has carrier. Running
+# this over wifi with no ethernet cable is not an image defect, so SKIP rather than
+# FAIL when eth0 has no carrier; plug a cable (or the direct laptop link) and re-run.
+if [ "$(cat /sys/class/net/eth0/carrier 2>/dev/null || echo 0)" = 1 ]; then
+	chk "eth0 holds 172.17.21.2"  "ip -4 addr show eth0 | grep -q '172.17.21.2'"
+else
+	skip "eth0 holds 172.17.21.2" "no ethernet carrier; plug a cable and re-run"
+fi
 
 echo "----"
-echo "$pass passed, $fail failed"
+echo "$pass passed, $fail failed, $skipped skipped"
 echo
 echo "Manual direct-cable check: set another machine's ethernet to 172.17.21.1,"
 echo "then from it:  ping 172.17.21.2   and   ssh <user>@172.17.21.2"