docs: de-tenant the manifest + home docstrings

Claude-Session: https://claude.ai/code/session_01Gq8suw7uuodAMPDRpo6iAB

middleware/studio_home.py

@@ -14,7 +14,7 @@ storage or lifecycle:
     POST /v1/library/fix         {path, instruction}  chat-fix: Qwen edit pass on that output
 
 One source of truth per concern, all pre-existing:
-  * assets + curation state  -> orgs/<org>/output/library.json (karma_manifest.py)
+  * assets + curation state  -> orgs/<org>/output/library.json (library_manifest.py)
   * what made an image       -> the API graph ComfyUI embeds in the PNG itself
   * running work             -> the ONE /prompt path (self-POST, so gpu_dispatch,
                                 billing and content_publish all apply unchanged)

scripts/library_manifest.py

@@ -2,7 +2,7 @@
 """library_manifest.py — regenerate the ONE canonical asset-library index.
 
 The Karma content pipeline has a single source of truth: a `library.json`
-manifest that sits at the root of the karma org's studio-output subtree and
+manifest that sits at the root of an org's studio-output subtree (any tenant; pass --root) and
 indexes every asset Antje has rendered. The storefront (karma.style) and the
 socials/blog queue both read THIS file — there is no second index.
 

studio_watchdog.sh

@@ -0,0 +1,47 @@
+#!/usr/bin/env bash
+# Studio watchdog: keep :8188 alive through the overnight karma render run.
+# Emits a line ONLY on a restart event (so it's a signal, not a firehose).
+# Detects: process gone, port unresponsive, or output-log silent >12min while a
+# render is supposedly in flight (the 6h-hang failure mode from the QC log).
+set -u
+STUDIO=/home/z/work/hanzo/studio
+LOG=$STUDIO/studio-local.log
+launch() {
+  cd "$STUDIO"
+  # Free the port first: a stale/duplicate main.py holding :8188 makes the new
+  # one crash on bind (EADDRINUSE). Kill any existing listener, wait for release.
+  pkill -f 'main.py --listen' 2>/dev/null || true
+  for _ in $(seq 1 15); do
+    ss -ltn 2>/dev/null | grep -q ':8188 ' || break
+    sleep 1
+  done
+  HF_HOME=/home/z/.cache/huggingface \
+  PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
+  STUDIO_PERSIST_QUEUE=1 \
+  nohup .venv/bin/python main.py --listen 0.0.0.0 --port 8188 --normalvram \
+    --disable-auto-launch --output-directory "$STUDIO/output" >> "$LOG" 2>&1 &
+  sleep 1
+  local pid
+  pid=$(pgrep -f 'main.py --listen' | head -1)
+  [ -n "$pid" ] && sudo -n sh -c "echo -700 > /proc/$pid/oom_score_adj" 2>/dev/null || true
+  echo "$pid"
+}
+while true; do
+  code=$(curl -s -m 5 http://127.0.0.1:8188/system_stats -o /dev/null -w "%{http_code}" 2>/dev/null || echo 000)
+  if [ "$code" != "200" ]; then
+    # give it one grace re-check (could be mid-render, momentarily busy)
+    sleep 8
+    code=$(curl -s -m 5 http://127.0.0.1:8188/system_stats -o /dev/null -w "%{http_code}" 2>/dev/null || echo 000)
+    if [ "$code" != "200" ]; then
+      pid=$(launch)
+      # wait for it to answer
+      for i in $(seq 1 45); do
+        c=$(curl -s -m 4 http://127.0.0.1:8188/system_stats -o /dev/null -w "%{http_code}" 2>/dev/null || echo 000)
+        [ "$c" = "200" ] && break
+        sleep 4
+      done
+      echo "STUDIO_RESTARTED pid=$pid health=$c"
+    fi
+  fi
+  sleep 45
+done