diff --git a/scripts/dispatcher.py b/scripts/dispatcher.py
index c2bb537..669f74c 100644
--- a/scripts/dispatcher.py
+++ b/scripts/dispatcher.py
@@ -105,6 +105,21 @@ def worker_free_vram_mib(worker: dict) -> int:
         return 0
 
 
+def _cleanup_stale_demo(worker: dict) -> bool:
+    """Kill the oldest demo.py on a worker with no active reconstruction to reclaim VRAM.
+    Returns True if a process was killed."""
+    rc, out, _ = ssh(worker["ssh_alias"], "pgrep -o -f 'python3.*demo\\.py' 2>/dev/null", timeout=10)
+    if rc != 0 or not out.strip():
+        return False
+    pid = out.strip().splitlines()[0].strip()
+    if not pid.isdigit():
+        return False
+    ssh(worker["ssh_alias"], f"kill {pid} 2>/dev/null", timeout=10)
+    print(f"  cleanup: killed stale demo.py pid={pid} on {worker['host']}", flush=True)
+    time.sleep(3)
+    return True
+
+
 _jobs_per_worker: dict[str, int] = {}
 
 
@@ -123,6 +138,10 @@ def pick_worker(estimated_vram_mib: int) -> dict | None:
             candidates.append(((load, -free, w["host"]), w))
         if not candidates:
             print(f"  pick_worker: no candidate for {estimated_vram_mib} MiB", flush=True)
+            # Free VRAM on idle workers by killing leftover demo.py (kept alive for viser).
+            for w in WORKERS:
+                if _jobs_per_worker.get(w["host"], 0) == 0:
+                    _cleanup_stale_demo(w)
             return None
         candidates.sort(key=lambda c: c[0])
         w = candidates[0][1]
@@ -323,6 +342,30 @@ def video_duration_s(worker: dict, worker_src: str) -> float:
         return 0.0
 
 
+def resolve_worker_video_source(worker: dict, video_path: str, frames_dir: str) -> tuple[str, bool]:
+    """Return a path readable by the worker and whether it is ephemeral cache.
+
+    Preferred path: direct read from the shared/external storage if the worker can see it.
+    Fallback: stage a src_*.MP4 copy inside the job frames_dir on the worker.
+    """
+    if ":" in video_path and not video_path.startswith("/"):
+        src_host, src_path = video_path.split(":", 1)
+        if src_host == worker["ssh_alias"] or src_host == worker["host"]:
+            return src_path, False
+    return f"{frames_dir}/src_{_path_basename(video_path)}", True
+
+
+def ensure_worker_video_source(video_path: str, worker: dict, worker_src: str, is_ephemeral: bool, step_label: str):
+    if not is_ephemeral:
+        return
+    rc_check = ssh(worker["ssh_alias"], f"test -f {shlex.quote(worker_src)}")[0]
+    if rc_check == 0:
+        return
+    print(f"  scp {_path_basename(video_path)} → {worker['host']}...")
+    set_status(step_label, step=f"scp: {_path_basename(video_path)}")
+    scp_to_worker(video_path, worker, worker_src)
+
+
 def do_extract(job: sqlite3.Row, worker: dict) -> str:
     videos = json.loads(job["video_paths"])
     frames_dir = f"{worker['frames_dir']}/job_{job['id']}"
@@ -333,15 +376,11 @@ def do_extract(job: sqlite3.Row, worker: dict) -> str:
     total_frames_est = 0  # will be computed after each scp
     total_duration_s = 0.0
     n_videos = len(videos)
-    for v in videos:
+    for vid_num, v in enumerate(videos, start=1):
         vf = f"fps={FPS},scale={IMG_W}:{IMG_H}"
         pattern = f"{frames_dir}/frame_%06d.jpg"
-        worker_src = f"{frames_dir}/src_{_path_basename(v)}"
-        rc_check = ssh(worker["ssh_alias"], f"test -f {shlex.quote(worker_src)}")[0]
-        if rc_check != 0:
-            print(f"  scp {_path_basename(v)} → {worker['host']}...")
-            set_status(job["id"], step=f"scp {idx // 1 + 1}/{n_videos}: {_path_basename(v)}")
-            scp_to_worker(v, worker, worker_src)
+        worker_src, is_ephemeral = resolve_worker_video_source(worker, v, frames_dir)
+        ensure_worker_video_source(v, worker, worker_src, is_ephemeral, job["id"])
         dur = video_duration_s(worker, worker_src)
         total_duration_s += dur
         total_frames_est += max(1, int(dur * FPS))
@@ -356,9 +395,6 @@ def do_extract(job: sqlite3.Row, worker: dict) -> str:
         )
         ssh(worker["ssh_alias"], f"setsid bash -c {shlex.quote(bg)} >/dev/null 2>&1 &")
 
-        # 1-based index for humans. We cannot compute it from `idx` directly because idx is
-        # the running frame counter, so count the loop iterations via total_duration_s order.
-        vid_num = videos.index(v) + 1
         thumb_refresh_counter = 0
         while True:
             # Use -s (file exists AND size > 0) to avoid race: setsid bash writes the exit code
@@ -383,9 +419,9 @@ def do_extract(job: sqlite3.Row, worker: dict) -> str:
             _, err, _ = ssh(worker["ssh_alias"], f"cat /tmp/cosma-ffmpeg-{job['id']}.log 2>/dev/null | tail -5 || echo ''")
             raise RuntimeError(f"ffmpeg failed on {v}: {err[:200]}")
         idx = count_frames(worker, frames_dir)
-        # Free MP4 cache immediately: thin pool on Proxmox host is tight and src_*.MP4
-        # are 1-11 GB each. Frames are already extracted so worker_src is no longer needed.
-        ssh(worker["ssh_alias"], f"rm -f {shlex.quote(worker_src)}")
+        # Free MP4 cache immediately only when we had to stage a local copy.
+        if is_ephemeral:
+            ssh(worker["ssh_alias"], f"rm -f {shlex.quote(worker_src)}")
         set_status(job["id"], frame_count=idx, progress=min(99, idx * 100 // total_frames_est))
     # Persist the measured video duration so the dashboard shows real length (segment_label
     # from ingest is only the timestamp of the first MP4 and lies when a segment spans multiple).
@@ -420,26 +456,27 @@ def do_reconstruct(job: sqlite3.Row, worker: dict, frames_dir: str) -> tuple[str
     log = f"/tmp/cosma-qc-job-{job['id']}.log"
     ckpt = f"{worker['lingbot_path']}/checkpoints/lingbot-map/lingbot-map-long.pt"
     ply_path = f"{frames_dir}/reconstruction.ply"
-    # Adaptive stride to fit CPU RAM: load_fn stacks full image tensor ~3.15 MB/frame @ 512x512x3 fp32.
-    # .87 has 23 GB RAM, .84 has 62 GB. Keep effective frame count ~4k to stay safe.
+    # More conservative RAM policy: recent runs die with rc=137 during image loading.
+    # Push stride earlier and keep windows smaller to trade speed for survival.
     frame_count = job["frame_count"] or 0
     ram_gb = 23 if worker["host"] == "192.168.0.87" else 62
-    ram_budget_gb = ram_gb * 0.35  # leave headroom for model + OS + cuda pinned buffers
+    ram_budget_gb = ram_gb * 0.22
     stride = 1
     while frame_count * 3.15 / 1024 / stride > ram_budget_gb:
         stride += 1
-    # demo.py starts a viser web server after saving the PLY and never exits.
-    # Wrap it: launch in bg, wait for "PLY saved" marker in the log, kill, exit 0.
-    # Match on the unique job frames_dir to identify our demo.py among all children/threads.
-    # Adapt window size to sequence length (lingbot-map README recommendation): bigger windows
-    # reduce overhead on long sequences. Effective frame count = frame_count / stride.
+    if frame_count > 4000:
+        stride = max(stride, 4)
+    elif frame_count > 2500:
+        stride = max(stride, 3)
+    elif frame_count > 1500:
+        stride = max(stride, 2)
     eff = frame_count // max(1, stride) if frame_count else 0
-    if eff > 3000:
-        window_size, overlap_size = 64, 16
-    elif eff > 320:
-        window_size, overlap_size = 32, 8
+    if eff > 2200:
+        window_size, overlap_size = 24, 6
+    elif eff > 900:
+        window_size, overlap_size = 16, 4
     else:
-        window_size, overlap_size = 16, 2
+        window_size, overlap_size = 12, 3
     marker = shlex.quote(frames_dir)
     cmd = (
         f"cd {shlex.quote(worker['lingbot_path'])} && source .venv/bin/activate && "