search: parse knowledgeGraph/peopleAlsoAsk; treat 0-result as ok+hint

Two robustness fixes for Serper search:
- Fold in knowledgeGraph (description + attributes) and peopleAlsoAsk
blocks. For many queries `organic` is empty but these carry the answer.
- When Serper genuinely returns nothing, return ok=True with an empty
result set + a "reformulate, drop quotes/rare acronyms" hint instead of
ok=False. Previously a 0-result Serper response was treated as a backend
failure, which fell through to the rate-limited DuckDuckGo path and
surfaced a misleading "all search backends failed" error. Now the agent
gets actionable feedback and the DDG 202 noise is gone. Mirrors
inference/tool_search.py "use a less specific query" behaviour.

app.py

@@ -1403,8 +1403,9 @@ def _serper_search(query: str, max_results: int) -> Dict[str, Any]:
                 "body": item.get("snippet", ""),
             }
         )
-    # Fold in the answer box and knowledge graph when present; these often
-    # carry the exact fact the model is looking for in a compact form.
+    # Fold in the answer box, knowledge graph and "people also ask" when
+    # present; for many queries `organic` is empty but these rich blocks
+    # still carry the exact fact the model is looking for.
     answer_box = data.get("answerBox") or {}
     if answer_box:
         rows.insert(
@@ -1417,13 +1418,50 @@ def _serper_search(query: str, max_results: int) -> Dict[str, Any]:
                 or "",
             },
         )
+    kg = data.get("knowledgeGraph") or {}
+    if kg:
+        kg_body = kg.get("description", "") or ""
+        attrs = kg.get("attributes") or {}
+        if attrs:
+            kg_body = (kg_body + "  " + "; ".join(
+                f"{k}: {v}" for k, v in attrs.items()
+            )).strip()
+        if kg.get("title") or kg_body:
+            rows.append(
+                {
+                    "title": kg.get("title", "Knowledge graph"),
+                    "href": kg.get("descriptionLink") or kg.get("website", ""),
+                    "body": kg_body,
+                }
+            )
+    for paa in (data.get("peopleAlsoAsk") or [])[:3]:
+        if paa.get("question"):
+            rows.append(
+                {
+                    "title": paa.get("question", ""),
+                    "href": paa.get("link", ""),
+                    "body": paa.get("snippet", ""),
+                }
+            )
     if not rows:
+        # Serper worked but the query genuinely matched nothing. Return a
+        # SUCCESSFUL-but-empty result (ok=True) carrying an actionable hint,
+        # rather than ok=False — that way the agent is told to reformulate
+        # instead of seeing a scary "all backends failed" message, and we
+        # skip the rate-limited DuckDuckGo fallback (which would also find
+        # nothing). Mirrors inference/tool_search.py's "use a less specific
+        # query" behaviour.
         return {
-            "ok": False,
+            "ok": True,
             "query": query,
-            "error": "Serper returned no organic results",
             "results": [],
+            "cached": False,
             "backend": "serper",
+            "note": (
+                "No results for this query. Reformulate and search again: "
+                "remove quotation marks (they force exact-phrase matching), "
+                "drop rare acronyms, and use fewer / more common keywords."
+            ),
         }
     return {
         "ok": True,