perf(movienet): vectorise eval matching; count frames missing from Image.zip

movienet_eval: replace the per-element dot() with numpy — actor references are
loaded once as an ndarray and scored with a single matmul, keeping a
whole-library gallery fast.

movienet_prep: count and report frames referenced by annotations but absent
from Image.zip instead of skipping them silently.

scripts/movienet_prep.py

@@ -133,20 +133,24 @@ def main():
     image_zip = movienet_root / "Image.zip"
     frame_cache: dict[str, np.ndarray] = {}
 
+    n_missing_in_zip = 0
     print(f"[prep] extracting {len(needed_paths)} frames from Image.zip…", file=sys.stderr)
     with zipfile.ZipFile(image_zip) as zf:
         for img_path in needed_paths:
             zip_entry = f"Image/{img_path}"
             try:
                 data = zf.read(zip_entry)
-                arr  = np.frombuffer(data, dtype=np.uint8)
-                img  = cv2.imdecode(arr, cv2.IMREAD_COLOR)
-                if img is not None:
-                    frame_cache[img_path] = img
             except KeyError:
-                pass  # file missing from zip, skip silently
+                n_missing_in_zip += 1  # frame referenced by an annotation but absent from Image.zip
+                continue
+            arr = np.frombuffer(data, dtype=np.uint8)
+            img = cv2.imdecode(arr, cv2.IMREAD_COLOR)
+            if img is not None:
+                frame_cache[img_path] = img
 
     print(f"[prep] frames loaded: {len(frame_cache)}/{len(needed_paths)}", file=sys.stderr)
+    if n_missing_in_zip:
+        print(f"[prep] frames absent from Image.zip: {n_missing_in_zip}", file=sys.stderr)
 
     per_actor_count: dict[str, int] = {}
     gt_entries = []