diff --git a/backend/Cargo.lock b/backend/Cargo.lock
index 940d03b..d55ff51 100644
--- a/backend/Cargo.lock
+++ b/backend/Cargo.lock
@@ -1517,7 +1517,7 @@ checksum = "c41e0c4fef86961ac6d6f8a82609f55f31b05e4fce149ac5710e439df7619ba4"
 
 [[package]]
 name = "mangalord"
-version = "0.87.7"
+version = "0.87.8"
 dependencies = [
  "anyhow",
  "argon2",
diff --git a/backend/Cargo.toml b/backend/Cargo.toml
index 65f9274..ee8219f 100644
--- a/backend/Cargo.toml
+++ b/backend/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "mangalord"
-version = "0.87.7"
+version = "0.87.8"
 edition = "2021"
 default-run = "mangalord"
 
diff --git a/frontend/package.json b/frontend/package.json
index 86a3bb1..8b3a37d 100644
--- a/frontend/package.json
+++ b/frontend/package.json
@@ -1,6 +1,6 @@
 {
   "name": "mangalord-frontend",
-  "version": "0.87.7",
+  "version": "0.87.8",
   "private": true,
   "type": "module",
   "scripts": {
diff --git a/vision-manager/manager.sh b/vision-manager/manager.sh
index 00dbd4d..e658abd 100644
--- a/vision-manager/manager.sh
+++ b/vision-manager/manager.sh
@@ -91,10 +91,27 @@ read_mem_used_pct() {
 }
 
 # ---- Docker helpers --------------------------------------------------------
-# `docker inspect .State.Running` → true/false; empty if the container is
-# absent (defined elsewhere and not yet created) — treated as not-running.
+# `docker inspect .State.Running` → true/false. Disambiguate "container
+# absent" (legitimate not-running) from "inspect failed" (transient
+# docker-socket-proxy hiccup) so the loop can skip a tick instead of
+# resetting timers as if vision had just been stopped — without this
+# distinction, an inspect blip every poll resets `up_for` and defeats
+# MAX_UPTIME entirely.
 vision_running() {
-  docker inspect -f '{{.State.Running}}' "$VISION_CONTAINER" 2>/dev/null || echo "false"
+  local out rc
+  out="$(docker inspect -f '{{.State.Running}}' "$VISION_CONTAINER" 2>&1)"
+  rc=$?
+  if [ "$rc" -eq 0 ]; then
+    # `true` or `false`. Echo as-is.
+    echo "$out"
+  elif echo "$out" | grep -qE 'No such container|No such object|not found'; then
+    # Container is intentionally not defined yet. Treat as not-running.
+    echo "false"
+  else
+    # Transient docker / socket-proxy error. Mirror the psql callers'
+    # `ERR` sentinel; the loop short-circuits the tick on this.
+    echo "ERR"
+  fi
 }
 
 vision_ready() {
@@ -111,7 +128,19 @@ start_vision() {
   local waited=0
   while ! vision_ready; do
     if [ "$waited" -ge "$START_HEALTH_TIMEOUT" ]; then
-      log "WARNING: $VISION_CONTAINER not healthy after ${START_HEALTH_TIMEOUT}s; leaving it running"
+      # Previously: log a warning and leave the container up. That made
+      # the wedge permanent — next loop tick sees running=true, the
+      # pending>0 branch falls through to `crawl_deferred=0` without
+      # retrying start, and `idle_for` never accumulates because
+      # `pending>0` resets it. The container stays running forever
+      # without ever answering /health.
+      #
+      # Stop it explicitly: a fresh `docker start` next tick takes the
+      # cold-load hit again but at least produces a clean state. If the
+      # wedge is persistent the operator sees a start/stop loop in the
+      # log instead of a silent hang.
+      log "WARNING: $VISION_CONTAINER not healthy after ${START_HEALTH_TIMEOUT}s; stopping so next tick can retry from a clean state"
+      stop_vision "unhealthy after ${START_HEALTH_TIMEOUT}s start probe" || true
       return 1
     fi
     sleep "$HEALTH_POLL_INTERVAL"
@@ -184,6 +213,16 @@ while true; do
       ;;
   esac
 
+  # Same treatment for the docker-inspect status. Without this, a transient
+  # docker-socket-proxy hiccup would mis-classify a running vision as stopped
+  # — every poll under load would reset `up_for`/`idle_for` and defeat
+  # MAX_UPTIME and the idle debounce both. Skip the tick and try again.
+  if [ "$running" = "ERR" ]; then
+    log "WARNING: docker inspect failed; will retry next tick"
+    sleep "$POLL_INTERVAL"
+    continue
+  fi
+
   # Gate on the runtime enable flag. If analysis is OFF, force the backlog to 0
   # so the normal idle path runs the STOP_DEBOUNCE timer and shuts vision down.
   # We deliberately reuse the idle debounce (rather than stopping instantly) so
@@ -238,7 +277,23 @@ while true; do
         # A RUNNING vision never reaches here; idle_for stayed 0 above because
         # pending>0, so a working vision is never idle-stopped under pressure.
         :
-      elif [ "$RESPECT_CRAWL_MUTEX" != "0" ] && [ "$(crawl_running)" != "0" ]; then
+      elif [ "$RESPECT_CRAWL_MUTEX" != "0" ] && {
+             cr="$(crawl_running)"
+             # Numeric guard for symmetry with analysis_enabled() (which
+             # fail-OPENs on ERR — see line 74). Without this, a psql
+             # blip falls through to the != "0" branch and we log the
+             # misleading "RAM mutex" line on every tick of the outage.
+             case "$cr" in
+               ''|*[!0-9]*)
+                 if [ "$crawl_deferred" != "1" ]; then
+                   log "WARNING: crawl_running query failed (got '${cr}') — deferring start until next clean read"
+                   crawl_deferred=1
+                 fi
+                 true
+                 ;;
+               *) [ "$cr" != "0" ] ;;
+             esac
+           }; then
         # Log once per deferral episode, not every poll, so a long crawl
         # doesn't flood the log.
         if [ "$crawl_deferred" != "1" ]; then