Use beat context for segmented action retiming

src/llm/vision_cache.py

@@ -283,7 +283,7 @@ def _semantic_match_score(beat_desc: str, candidate_desc: str) -> tuple[float, s
     beat_actions = _semantic_action_groups(beat_desc)
     candidate_actions = _semantic_action_groups(candidate_desc)
     required = beat_actions & _STRONG_ACTION_GROUPS
-    missing = required - candidate_actions
+    missing = _missing_action_groups(required, candidate_actions)
     if missing:
         penalty = min(0.45, 0.18 * len(missing))
         text_score = max(0.0, text_score - penalty)
@@ -297,6 +297,80 @@ def _semantic_match_score(beat_desc: str, candidate_desc: str) -> tuple[float, s
     return text_score, reason
 
 
+def _missing_action_groups(required: set[str], candidate_actions: set[str]) -> set[str]:
+    missing = set(required) - set(candidate_actions)
+    if "approach" in missing and ({"kiss", "forehead_touch"} & set(candidate_actions)):
+        missing.remove("approach")
+    return missing
+
+
+def _action_phase_text(desc: str) -> str:
+    match = re.search(r'"action_phase"\s*:\s*"([^"]+)"', desc, flags=re.IGNORECASE | re.DOTALL)
+    if match:
+        return match.group(1).lower()
+    return desc.lower()
+
+
+def _action_phase_adjustment(beat_desc: str, candidate_desc: str) -> tuple[float, str]:
+    """
+    Score whether a candidate is in the same action phase, not just the same scene.
+
+    Vision text can make two moments in one long dialogue sound equally similar:
+    "moving closer until foreheads touch" and "static conversation after contact"
+    share subjects/composition, but their temporal phase is different. This
+    adjustment is intentionally generic and only looks for broad phase words.
+    """
+    beat_phase = _action_phase_text(beat_desc)
+    candidate_phase = _action_phase_text(candidate_desc)
+    adjustment = 0.0
+    notes: list[str] = []
+
+    beat_dynamic = any(term in beat_phase for term in (
+        "moving", "move", "approach", "approaching", "lean", "leaning",
+        "raise", "raising", "lift", "lifting", "turn", "turning",
+        "pull", "pulling", "transition",
+    ))
+    candidate_static = any(term in candidate_phase for term in (
+        "static", "conversation", "talking", "speaking", "listening",
+        "subtle facial", "slight facial",
+    ))
+    if beat_dynamic and candidate_static:
+        adjustment -= 0.18
+        notes.append("dynamic_vs_static")
+
+    beat_contact_target = any(term in beat_phase for term in (
+        "forehead", "foreheads", "touch", "kiss",
+    ))
+    candidate_contact = any(term in candidate_phase for term in (
+        "forehead", "foreheads", "touch", "touching", "kiss", "kissing",
+    ))
+    candidate_after_contact = any(term in candidate_phase for term in (
+        "gap", "pull back", "pulling back", "look at each other",
+        "looks at each other", "conversation",
+    ))
+    if beat_contact_target and candidate_contact:
+        adjustment += 0.08
+        notes.append("contact")
+    if beat_contact_target and candidate_after_contact and "until" in beat_phase:
+        adjustment -= 0.28
+        notes.append("after_contact")
+
+    beat_kiss = "kiss" in beat_phase or "kissing" in beat_phase
+    candidate_kiss = "kiss" in candidate_phase or "kissing" in candidate_phase
+    if beat_kiss and candidate_kiss:
+        adjustment += 0.12
+        notes.append("kiss")
+    elif beat_contact_target and candidate_kiss:
+        adjustment += 0.06
+        notes.append("kiss_contact")
+
+    if "talk" not in beat_phase and "conversation" in candidate_phase:
+        adjustment -= 0.10
+        notes.append("unexpected_conversation")
+
+    return adjustment, ",".join(notes) if notes else "phase_neutral"
+
+
 def _scene_seed_points(scene: Scene, max_points: int) -> list[float]:
     if max_points <= 1 or scene.duration_s <= 0:
         return [scene.start_s]
@@ -612,7 +686,7 @@ def validate_match_window_with_vision(
     score, reason = _semantic_match_score(beat_desc, source_desc)
     beat_actions = _semantic_action_groups(beat_desc) & _STRONG_ACTION_GROUPS
     source_actions = _semantic_action_groups(source_desc)
-    missing_actions = beat_actions - source_actions
+    missing_actions = _missing_action_groups(beat_actions, source_actions)
     threshold = max(0.32, cfg.vision.similarity_threshold + 0.12)
     if missing_actions and score < threshold:
         return False, f"{reason} missing_actions={sorted(missing_actions)}"
@@ -675,14 +749,25 @@ def find_action_window_in_scene(
             continue
         score, reason = _semantic_match_score(beat_desc, desc)
         source_actions = _semantic_action_groups(desc)
-        missing_actions = beat_actions - source_actions
+        missing_actions = _missing_action_groups(beat_actions, source_actions)
         if missing_actions:
             continue
         threshold = max(0.38, cfg.vision.similarity_threshold + 0.18)
         if score < threshold:
             continue
-        candidate = (start_s, end_s, score, reason)
-        if best is None or candidate[2] > best[2]:
+        phase_adjustment, phase_reason = _action_phase_adjustment(beat_desc, desc)
+        adjusted_score = max(0.0, min(1.0, score + phase_adjustment))
+        if adjusted_score < threshold:
+            continue
+        candidate = (
+            start_s,
+            end_s,
+            adjusted_score,
+            f"{reason} phase={phase_reason} raw={score:.3f}",
+        )
+        if best is None or candidate[2] > best[2] + 0.03 or (
+            candidate[2] >= best[2] - 0.03 and candidate[0] < best[0]
+        ):
             best = candidate
 
     _save_cache(cfg, cache)