chore: config — primary를 gemma4:31b-cloud로 정정 + 법령 API 키 배선

- provider ollama_local→ollama, primary/roles를 gemma4:26b→gemma4:31b-cloud로
  (메모리 기준선과 일치, 런타임이 이미 이 값으로 운영 중이던 것을 커밋에 반영)
- gemma4:26b(지서버) / muse-glimmer 프로필에 fixedNumCtx, 26b엔 airkorea 안내 추가
- mistral-large-3 프로필에서 중복 forceToolChoiceOnLiveData 제거(전역 ON됨)
- legal.law_go_kr_api_key vault 참조 추가

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HdchNLTHsyYahp2PVQjoym

.smallclaw/config.json

@@ -17,11 +17,11 @@
     }
   },
   "llm": {
-    "provider": "ollama_local",
+    "provider": "ollama",
     "providers": {
       "ollama": {
         "endpoint": "http://localhost:11434",
-        "model": "gemma4:26b"
+        "model": "gemma4:31b-cloud"
       },
       "lm_studio": {
         "endpoint": "http://192.168.0.7:1234",
@@ -56,21 +56,27 @@
     }
   },
   "models": {
-    "primary": "gemma4:26b",
+    "primary": "gemma4:31b-cloud",
     "fallback": "kimi-k2.6:cloud",
     "roles": {
-      "manager": "gemma4:26b",
-      "executor": "gemma4:26b",
-      "verifier": "gemma4:26b",
+      "manager": "gemma4:31b-cloud",
+      "executor": "gemma4:31b-cloud",
+      "verifier": "gemma4:31b-cloud",
       "background_task": ""
     },
     "profiles": {
       "mistral-large-3:675b-cloud": {
-        "extraSystemPrompt": "You have a documented tendency to skip tool calls and answer fluently and confidently from memory instead — on news, weather, and factual/product questions (specs, prices, versions, comparisons) alike, sometimes inventing specific-sounding numbers or even nonexistent events. Before answering ANYTHING with a checkable real-world fact, call the relevant tool (web_search, weather_kma, etc.) first — do not trust your own confidence as a substitute for checking. If a tool call fails or returns nothing useful, say so plainly instead of filling the gap from memory.",
-        "forceToolChoiceOnLiveData": true
+        "extraSystemPrompt": "You have a documented tendency to skip tool calls and answer fluently and confidently from memory instead — on news, weather, and factual/product questions (specs, prices, versions, comparisons) alike, sometimes inventing specific-sounding numbers or even nonexistent events. Before answering ANYTHING with a checkable real-world fact, call the relevant tool (web_search, weather_kma, etc.) first — do not trust your own confidence as a substitute for checking. If a tool call fails or returns nothing useful, say so plainly instead of filling the gap from memory."
       },
       "kimi-k2.6:cloud": {
         "extraSystemPrompt": "You have a documented tendency to massively over-search — observed doing 7+ (once 50+) web_search/web_fetch calls for a single broad request like \"오늘 국제 뉴스 정리\", searching topic-by-topic (Ukraine, Gaza, tariffs, ...) instead of a couple of broad searches. A hard cap now blocks you past 5 web_search/web_fetch calls per turn — but don't rely on the cap. Plan your search angles up front, pick the 1-3 most important ones, and write the answer once you have enough. Exhaustive topic-by-topic coverage is not the goal."
+      },
+      "gemma4:26b": {
+        "extraSystemPrompt": "You are running on 지서버 as gemma4:26b (via Ollama). weather_kma carries no air-quality data: for 미세먼지/PM2.5/PM10/황사 call weather_airkorea instead — re-calling weather_kma with the same arguments will never return air quality.",
+        "fixedNumCtx": 262144
+      },
+      "muse-glimmer:latest": {
+        "fixedNumCtx": 131072
       }
     }
   },
@@ -386,5 +392,8 @@
   "collabora": {
     "baseUrl": "https://col.applecherry.net",
     "publicUrl": "https://ai.applecherry.net"
+  },
+  "legal": {
+    "law_go_kr_api_key": "vault:legal.law_go_kr_api_key"
   }
 }