ingest+dispatcher — support acquisition depuis remote host via SSH

- ingest.py : --remote-host <alias> pour scanner/exiftool via SSH, stocke
  les chemins avec préfixe "alias:" pour que le worker sache puller direct
- dispatcher.py : scp_to_worker détecte "host:path" et fait pull remote
  (worker → source host) au lieu du double hop via dispatcher
- _path_basename gère les paths préfixés pour ffmpeg

Permet d'ingester les vidéos depuis n'importe quelle machine accessible
en SSH sans passer 145GB par le conteneur FastAPI.

scripts/dispatcher.py

@@ -121,6 +121,24 @@ def count_frames(worker: dict, frames_dir: str) -> int:
 
 
 def scp_to_worker(local_path: str, worker: dict, remote_path: str):
+    """Copy a file to the worker.
+
+    `local_path` may be either:
+      - a path on the dispatcher host (standard scp from here)
+      - "host:abs_path" — pulled by the worker directly from `host`
+        (avoids routing bytes through the dispatcher).
+    """
+    if ":" in local_path and not local_path.startswith("/"):
+        src_host, src_path = local_path.split(":", 1)
+        # Pull from source host directly on the worker
+        pull_cmd = (
+            f"scp -o BatchMode=yes {shlex.quote(src_host)}:{shlex.quote(src_path)} "
+            f"{shlex.quote(remote_path)}"
+        )
+        rc, _, err = ssh(worker["ssh_alias"], pull_cmd, timeout=7200)
+        if rc != 0:
+            raise RuntimeError(f"remote scp ({src_host}→{worker['host']}) failed: {err[:200]}")
+        return
     r = subprocess.run(
         ["scp", "-o", "BatchMode=yes", local_path, f"{worker['ssh_alias']}:{remote_path}"],
         capture_output=True, timeout=1800,
@@ -129,6 +147,12 @@ def scp_to_worker(local_path: str, worker: dict, remote_path: str):
         raise RuntimeError(f"scp failed: {r.stderr.decode()[:200]}")
 
 
+def _path_basename(p: str) -> str:
+    if ":" in p and not p.startswith("/"):
+        return Path(p.split(":", 1)[1]).name
+    return Path(p).name
+
+
 def do_extract(job: sqlite3.Row, worker: dict) -> str:
     videos = json.loads(job["video_paths"])
     frames_dir = f"{worker['frames_dir']}/job_{job['id']}"
@@ -138,10 +162,10 @@ def do_extract(job: sqlite3.Row, worker: dict) -> str:
         vf = f"fps={FPS},scale={IMG_W}:{IMG_H}"
         pattern = f"{frames_dir}/frame_%06d.jpg"
         # Copy video to worker if it doesn't exist there
-        worker_src = f"{frames_dir}/src_{Path(v).name}"
+        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(v).name} → {worker['host']}...")
+            print(f"  scp {_path_basename(v)} → {worker['host']}...")
             scp_to_worker(v, worker, worker_src)
         cmd = (
             f"ffmpeg -hide_banner -loglevel error -i {shlex.quote(worker_src)} "