perf(search): offload cosine scoring off event loop; document best-effort feed

Drift-audit Group 9 (param-cliff / unbounded search work):

- semantic_search_notes: the O(rows) cosine-similarity scoring loop ran
  synchronously on the event loop, so every RAG injection / search stalled
  other requests proportional to the user's embedding count. Move the scoring
  into asyncio.to_thread (results unchanged). The deeper fix — bounding the
  candidate set via pgvector ORDER BY/LIMIT — is noted as separate infra work.
- _semantic_knowledge_search: documented the best-effort top-N semantics —
   is the capped candidate-window size (not the true match count),
  matches beyond the cap aren't page-reachable, and each page recomputes the
  full merge. Prevents the silent-truncation trap; cached ranked-id paging /
  pgvector is the fix if exhaustive pagination is ever required.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

src/fabledassistant/services/embeddings.py

@@ -153,17 +153,22 @@ async def semantic_search_notes(
     if not rows:
         return []
 
-    scored: list[tuple[float, Note]] = []
-    for ne, note in rows:
-        try:
-            sim = _cosine_similarity(query_vec, ne.embedding)
-        except Exception:
-            continue
-        if sim >= threshold:
-            scored.append((sim, note))
+    def _score() -> list[tuple[float, Note]]:
+        out: list[tuple[float, Note]] = []
+        for ne, note in rows:
+            try:
+                sim = _cosine_similarity(query_vec, ne.embedding)
+            except Exception:
+                continue
+            if sim >= threshold:
+                out.append((sim, note))
+        out.sort(key=lambda x: x[0], reverse=True)
+        return out[:limit]
 
-    scored.sort(key=lambda x: x[0], reverse=True)
-    return scored[:limit]
+    # Offload the O(rows) cosine scoring off the event loop so a large corpus
+    # doesn't stall other requests while ranking. Results are unchanged; the
+    # real scaling fix (ORDER BY / LIMIT in pgvector) is a separate effort.
+    return await asyncio.to_thread(_score)
 
 
 async def backfill_note_embeddings() -> None: