fix: skill search ranking - use overview for embedding and fix visited set filtering

1. skill_processor.py: Use LLM-generated overview for vectorization
   instead of short abstract, aligning with how resources handle
   directory vectorization in semantic_processor.py.

2. hierarchical_retriever.py: Separate 'visited for traversal' from
   'collected as result'. The visited set previously dropped the most
   relevant results - global search found them first, marked visited,
   then parent directory search skipped them as children.

Author: ZaynJarvis

openviking/retrieve/hierarchical_retriever.py

@@ -320,20 +320,27 @@ def merge_filter(base_filter: Dict, extra_filter: Optional[Dict]) -> Dict:
                     alpha * score + (1 - alpha) * current_score if current_score else score
                 )
 
-                if passes_threshold(final_score) and uri not in visited:
+                if not passes_threshold(final_score):
+                    logger.debug(
+                        f"[RecursiveSearch] URI {uri} score {final_score} did not pass threshold {effective_threshold}"
+                    )
+                    continue
+
+                # Always collect results that pass threshold, even if already
+                # visited as a directory starting point. The visited set only
+                # prevents re-entering directories for child search.
+                if not any(c.get("uri") == uri for c in collected):
                     r["_final_score"] = final_score
                     collected.append(r)
                     logger.debug(
                         f"[RecursiveSearch] Added URI: {uri} to candidates with score: {final_score}"
                     )
+
+                if uri not in visited:
                     if r.get("is_leaf"):
                         visited.add(uri)
-                        continue
-                    heapq.heappush(dir_queue, (-final_score, uri))
-                else:
-                    logger.debug(
-                        f"[RecursiveSearch] URI {uri} score {final_score} did not pass threshold {effective_threshold}"
-                    )
+                    else:
+                        heapq.heappush(dir_queue, (-final_score, uri))
 
             # Convergence check
             current_topk = sorted(collected, key=lambda x: x.get("_final_score", 0), reverse=True)[

openviking/utils/skill_processor.py

@@ -74,10 +74,12 @@ async def process_skill(
                 "source_path": skill_dict.get("source_path", ""),
             },
         )
-        context.set_vectorize(Vectorize(text=context.abstract))
-
         overview = await self._generate_overview(skill_dict, config)
 
+        # Use overview for vectorization (richer semantic content than abstract alone)
+        vectorize_text = overview if overview else context.abstract
+        context.set_vectorize(Vectorize(text=vectorize_text))
+
         skill_dir_uri = f"viking://agent/skills/{context.meta['name']}"
 
         await self._write_skill_content(