diff --git a/tests/lifecycle/bats/immich.bats b/tests/lifecycle/bats/immich.bats
index b3779875..fec60fab 100644
--- a/tests/lifecycle/bats/immich.bats
+++ b/tests/lifecycle/bats/immich.bats
@@ -47,9 +47,23 @@ teardown_file() {
 }
 
 @test "immich exposes its web UI lan-address (port 2283)" {
-  run rpc_result container-list
-  [ "$status" -eq 0 ]
-  echo "$output" | jq -e '.[] | select(.name == "immich") | .lan_address | test("2283")' >/dev/null
+  # Poll briefly: lan_address is derived from the published host port, which is
+  # momentarily absent (null) while immich_server is mid-recreate (e.g. a
+  # health-monitor bounce during the read-only tier). A genuinely unexposed
+  # immich never publishes 2283, so this still catches real port drift; it only
+  # absorbs the transient null seen under churn.
+  local deadline=$(( $(date +%s) + 30 ))
+  while (( $(date +%s) < deadline )); do
+    run rpc_result container-list
+    [ "$status" -eq 0 ]
+    if echo "$output" \
+      | jq -e '.[] | select(.name == "immich") | .lan_address // "" | test("2283")' >/dev/null; then
+      return 0
+    fi
+    sleep 3
+  done
+  echo "immich never reported a lan_address containing 2283 within 30s" >&2
+  return 1
 }
 
 # ────────────────────────────────────────────────────────────────────