Use motion phase for in-scene timing

cli.py

@@ -640,7 +640,7 @@ def _filter_semantically_invalid_vision_matches(results: list, beats: list, cfg)
     from dataclasses import replace
     from src.llm.vision_cache import find_action_window_in_scene, validate_match_window_with_vision
     from src.cv.scene_indexer import build_scene_index
-    from src.cv.global_scan import align_in_point_by_content
+    from src.cv.global_scan import align_in_point_by_content, align_in_point_by_motion
 
     logger = logging.getLogger(__name__)
     beats_by_id = {beat.beat_id: beat for beat in beats}
@@ -655,12 +655,18 @@ def _filter_semantically_invalid_vision_matches(results: list, beats: list, cfg)
             return None
         start_s, end_s, semantic_score, reason = found
         window_s = max(1.0, min(4.0, (end_s - start_s) * 1.5))
-        aligned_in_s, content_score = align_in_point_by_content(
+        motion_in_s, motion_score = align_in_point_by_motion(
             check_beat,
             start_s,
             cfg,
             search_window_s=window_s,
         )
+        aligned_in_s, content_score = align_in_point_by_content(
+            check_beat,
+            motion_in_s,
+            cfg,
+            search_window_s=min(window_s, 0.8),
+        )
         aligned_in_s = max(scene.start_s, min(aligned_in_s, max(scene.start_s, scene.end_s - check_beat.duration_s)))
         ok, verify_reason = validate_match_window_with_vision(
             check_beat,
@@ -679,7 +685,7 @@ def _filter_semantically_invalid_vision_matches(results: list, beats: list, cfg)
                 verify_reason,
             )
             return None
-        score = max(content_score, min(0.99, semantic_score))
+        score = max(content_score, min(0.99, semantic_score * 0.75 + motion_score * 0.25))
         return scene, aligned_in_s, score, f"{reason}; {verify_reason}"
 
     kept = []