dispatcher — keep demo.py alive apres PLY + auto-clear error

1. Ne plus kill demo.py apres PLY saved: son viser/PointCloudViewer natif
   (camera frustums, per-frame confidence filtering, animation) donne une
   visu bcp plus propre que viser_ply.py standalone (XYZ+RGB seul).
   Cout: ~6GB VRAM par job done garde alive jusquau prochain pick_worker
   qui peut kill si besoin.

2. set_status clear auto le champ error quand status transitionne vers
   extracting/running/done/queued: sinon les dashboards montrent les
   erreurs historiques sur les jobs en cours de retry.

docker-compose.yml

@@ -6,13 +6,11 @@ services:
     ports:
       - "3849:8000"
     volumes:
-      - cosma-qc-db:/var/lib/cosma-qc
+      - /home/floppyrj45/cosma-qc-data:/var/lib/cosma-qc
+      - /home/floppyrj45/.ssh:/ssh-in:ro
     environment:
       COSMA_QC_WORKERS: |
         [
-          {"host":"192.168.0.87","ssh_alias":"gpu","gpu":"RTX 3060 12GB"},
-          {"host":"192.168.0.84","ssh_alias":"ml-stack","gpu":"RTX 3090 24GB"}
+          {"host":"192.168.0.87","ssh_alias":"gpu","gpu":"RTX 3060 12GB","vram_mib":11913,"frames_dir":"/home/floppyrj45/cosma-qc-frames","lingbot_path":"/home/floppyrj45/ai-video/lingbot-map","viser_port_base":8100},
+          {"host":"192.168.0.84","ssh_alias":"ml-stack","gpu":"RTX 3090 24GB","vram_mib":24576,"frames_dir":"/root/cosma-qc-frames","lingbot_path":"/root/ai-video/lingbot-map","viser_port_base":8100}
         ]
-
-volumes:
-  cosma-qc-db:

scripts/dispatcher.py

@@ -115,6 +115,10 @@ def estimate_vram_mib(frame_count: int) -> int:
 
 
 def set_status(job_id: int, **fields):
+    # Auto-clear stale error text when the job moves into a live state so the dashboard
+    # stops showing a previous failure alongside a fresh run.
+    if fields.get("status") in ("extracting", "running", "done", "queued") and "error" not in fields:
+        fields["error"] = None
     keys = list(fields.keys())
     vals = [fields[k] for k in keys]
     q = "UPDATE jobs SET " + ", ".join(f"{k}=?" for k in keys) + " WHERE id=?"
@@ -263,10 +267,11 @@ def do_reconstruct(job: sqlite3.Row, worker: dict, frames_dir: str) -> tuple[str
         f"for i in $(seq 1 3600); do "
         f"  if ! kill -0 $DEMO_PID 2>/dev/null; then wait $DEMO_PID; exit $?; fi; "
         f"  if grep -q 'PLY saved:' {log} 2>/dev/null; then "
-        f"    sleep 2; "
-        f"    pkill -TERM -f \"demo.py.*{frames_dir}\" 2>/dev/null; sleep 1; "
-        f"    pkill -KILL -f \"demo.py.*{frames_dir}\" 2>/dev/null; "
-        f"    wait $DEMO_PID 2>/dev/null; exit 0; "
+        # Keep demo.py alive so its viser/PointCloudViewer (with camera frustums, per-frame
+        # confidence filtering, animation) stays reachable. Standalone viser_ply.py only has
+        # XYZ+RGB which gives a poor-looking cloud. The worker eats ~6GB VRAM per alive demo.py
+        # until pick_worker can no longer fit a new job; _cleanup_stale_demos reaps the oldest.
+        f"    exit 0; "
         f"  fi; "
         f"  sleep 3; "
         f"done; "