feat/igpu-ov: --probe-stride/--probe-max-per-scene — thin inference (not decode) on slow devices

Every-frame prediction is free on CUDA (~0.2s) and hopeless on the iGPU
(5.2s/img): a 100-frame faceless scene cost 8.7 min. stride>1 keeps
frame-perfect decode and probes every Nth frame, max M per scene, then
waits for the next scene start. Default stride=1 = original behavior.
Scheduler unit-tested (face-hit, faceless-exhaust, multi-scene).

anime_facesaver.py

@@ -119,6 +119,22 @@ def parse_arguments():
              "edges. 0 (default) allows cropped faces; set >0 to reject "
              "faces touching the crop lines.",
     )
+    parser.add_argument(
+        "--probe-stride",
+        type=int,
+        default=1,
+        help="Within a scene still needing a face, run detection every Nth "
+             "frame instead of every frame (1 = original behavior, free on "
+             "CUDA). For slow devices (iGPU ~5s/img) use e.g. 25: decode "
+             "stays frame-perfect, only the expensive inference is thinned.",
+    )
+    parser.add_argument(
+        "--probe-max-per-scene",
+        type=int,
+        default=4,
+        help="With --probe-stride>1, probes before giving up on a scene and "
+             "waiting for the next scene start (default 4).",
+    )
     parser.add_argument(
         "--scene-threshold",
         type=float,
@@ -276,6 +292,17 @@ def process_video(video_path, output_dir, args, model, device, draw,
     face_in_scene = False
     frame_idx = 0
 
+    # Probe schedule. probe_stride=1 reproduces the original every-frame
+    # loop (right on CUDA, where predict costs ~0.2 s). On slow devices
+    # (iGPU ~5 s/img) predicting every frame is hopeless: decode is ~300 fps,
+    # inference 0.2/s. With stride>1 we decode through frames for free and
+    # predict only at scene starts, retrying every stride frames up to
+    # probe_max attempts per scene, then waiting for the next scene start.
+    probe_stride = getattr(args, "probe_stride", 1)
+    probe_max = getattr(args, "probe_max_per_scene", 4)
+    next_probe = 0
+    probes_used = 0
+
     while True:
         ret, frame = cap.read()
         if not ret:
@@ -283,10 +310,20 @@ def process_video(video_path, output_dir, args, model, device, draw,
 
         if frame_idx in scene_starts:
             face_in_scene = False
-        # membership test on a set; pop matched starts to keep it cheap
-        # (scenes strictly increase, so set stays small either way)
+            next_probe = frame_idx
+            probes_used = 0
+
+        if face_in_scene:
+            frame_idx += 1
+            continue
+        if probe_stride > 1 and (frame_idx < next_probe
+                                 or probes_used >= probe_max):
+            frame_idx += 1
+            continue
 
         if not face_in_scene:
+            probes_used += 1
+            next_probe = frame_idx + probe_stride
             try:
                 results = model.predict(frame, conf=args.conf,
                                         iou=args.iou, imgsz=args.imgsz,