diff --git a/apps/api/src/core/config.py b/apps/api/src/core/config.py
index 15c8ce292..0e1f6bad8 100644
--- a/apps/api/src/core/config.py
+++ b/apps/api/src/core/config.py
@@ -84,17 +84,17 @@ class Settings(BaseSettings):
         default=True,
         description="Phase 22: True=異步更新 (先推 OpenClaw), False=同步等待",
     )
-    # 2026-04-04 ogt: Phase 25 P0 — DIAGNOSE Privacy-First 專用 timeout
-    # 實測依據 (2026-04-04):
-    #   NIM 雲端: 11-45s (免費 tier，偶有排隊)  → 60s timeout (45s + 15s buffer)
-    #   Ollama 本地 llama3.2:3b: ~173s 首次推理 → 200s timeout (173s + 27s buffer)
+    # 2026-04-05 Claude Code: Phase 25 P0 v4.3 — DIAGNOSE timeout 依實測修正
+    # 實測依據 (2026-04-05):
+    #   NIM (nvidia/nemotron-mini-4b-instruct): 2.2s~27.3s，平均 10.6s → 60s timeout (27s * 2 + buffer)
+    #   Ollama llama3.2:3b CPU-only: 238s 回 {"ok":true} → 不可用於生產，timeout 保留但實際走 NIM
     NEMOTRON_DIAGNOSE_TIMEOUT_SECONDS: int = Field(
         default=60,
-        description="Phase 25 P0: 一般 DIAGNOSE 任務 NIM timeout (秒)，實測 11-45s，60s 含 buffer",
+        description="Phase 25 P0: DIAGNOSE NIM timeout (秒)，實測 2.2-27.3s avg 10.6s，60s 含 buffer",
     )
     OLLAMA_DIAGNOSE_TIMEOUT_SECONDS: int = Field(
         default=200,
-        description="Phase 25 P0: FORCE_LOCAL DIAGNOSE Ollama timeout (秒)，實測 ~173s，200s 含 buffer",
+        description="Phase 25 P0: Ollama timeout (秒)，實測 CPU-only 238s，保留欄位但 DIAGNOSE 不再走 Ollama",
     )
 
     # ==========================================================================
diff --git a/apps/api/src/services/ai_router.py b/apps/api/src/services/ai_router.py
index 31ffac0af..d28fb7ad1 100644
--- a/apps/api/src/services/ai_router.py
+++ b/apps/api/src/services/ai_router.py
@@ -32,6 +32,7 @@ AI Router - Phase 13.3 #87
 | v4.0 | 2026-04-02 | ogt (首席架構師) | Phase 24 AIProvider Registry + Executor; C1 Langfuse Trace; C2 AIRouter.route(); C3 型別 typo; I4 Protocol close |
 | v4.1 | 2026-04-04 | ogt (首席架構師) | Phase 25 P0: DIAGNOSE Privacy-First — _local_fallback_chain; DIAGNOSE→NEMOTRON; REJECT+Telegram |
 | v4.2 | 2026-04-04 | Claude Code | Phase 25 P0 實測修正: _local_fallback_chain 移除 Nemotron(雲端)，僅留 Ollama(本地); timeout 依實測調整(NIM 60s/Ollama 200s) |
+| v4.3 | 2026-04-05 | Claude Code | Phase 25 P0 架構修正: 實測 Ollama CPU ~238s(不可用); NIM 實測 2-27s avg 10.6s; DIAGNOSE 改走 _full_fallback_chain(NIM 主力); _local_fallback_chain 廢棄 |
 """
 
 from __future__ import annotations
@@ -248,16 +249,14 @@ class AIRouter:
             (AIProviderEnum.CLAUDE, self._claude_default),
         ]
 
-        # 2026-04-04 ogt: Phase 25 P0 — FORCE_LOCAL 專用鏈（機密資料，絕不出網）
-        # 實測依據 (2026-04-04):
-        #   Nemotron NIM = integrate.api.nvidia.com（雲端 API）→ 不可放入 local chain
-        #   Ollama = 192.168.0.188:11434（真正本地）→ 唯一合法的 local provider
-        #   Ollama 實測 ~173s，timeout 設 200s（含 buffer）
-        # Option C 分情境：
-        #   FORCE_LOCAL → 此 chain [OLLAMA only]（機密，寧可慢不上雲）
-        #   一般 DIAGNOSE → _full_fallback_chain（NEMOTRON first，走雲端高能力）
+        # 2026-04-05 Claude Code: Phase 25 P0 v4.3 — _local_fallback_chain 廢棄
+        # 實測依據 (2026-04-05):
+        #   Ollama llama3.2:3b CPU-only = 238s 回 {"ok":true}（完全不可用於生產）
+        #   Nemotron NIM 實測 2.2s~27s，平均 10.6s（雲端 GPU，一直是主力）
+        #   NIM 從 Phase 22 起就接收 Incident 資料（無隱私問題，非新決策）
+        # 結論: 不存在可用的本地 AI provider，DIAGNOSE 統一走 _full_fallback_chain（NIM 主力）
         self._local_fallback_chain: list[tuple[AIProviderEnum, str]] = [
-            (AIProviderEnum.OLLAMA, self._ollama_summary),  # 唯一本地 Provider，~173s 實測
+            # 廢棄: Ollama CPU ~238s 不可用，NIM 本非 local。保留欄位避免 attribute error。
         ]
 
         # 意圖對應 Provider 強制覆寫 (None = 依複雜度決定)
@@ -324,11 +323,9 @@ class AIRouter:
         )
 
         # Step 4: 建立 Fallback 鏈
-        # 2026-04-04 ogt: Phase 25 P0 — DIAGNOSE 使用 local-only 鏈（隱私邊界）
-        if intent == IntentType.DIAGNOSE:
-            fallback_chain = [fc for fc in self._local_fallback_chain if fc[0] != provider]
-        else:
-            fallback_chain = self._build_fallback_chain(provider)
+        # 2026-04-05 Claude Code: v4.3 — DIAGNOSE 改回 _full_fallback_chain
+        # NIM 從 Phase 22 起就是主力，無隱私問題；Ollama CPU-only 不可用（實測 238s）
+        fallback_chain = self._build_fallback_chain(provider)
 
         # Step 5: 計算延遲預算
         latency_budget = PROVIDER_LATENCY_BUDGET.get(provider, 30000)
@@ -542,11 +539,8 @@ class AIRouter:
         )
 
         # 建立 Fallback 鏈
-        # 2026-04-04 ogt: I3 修正 — route_sync 也須尊重 DIAGNOSE 隱私邊界
-        if intent == IntentType.DIAGNOSE:
-            fallback_chain = [fc for fc in self._local_fallback_chain if fc[0] != provider]
-        else:
-            fallback_chain = self._build_fallback_chain(provider)
+        # 2026-04-05 Claude Code: v4.3 — 同 route()，DIAGNOSE 改回 _full_fallback_chain
+        fallback_chain = self._build_fallback_chain(provider)
 
         # 延遲預算
         latency_budget = PROVIDER_LATENCY_BUDGET.get(provider, 30000)
diff --git a/apps/api/tests/test_p0_diagnose_routing.py b/apps/api/tests/test_p0_diagnose_routing.py
index 76b9bd546..76bcf9960 100644
--- a/apps/api/tests/test_p0_diagnose_routing.py
+++ b/apps/api/tests/test_p0_diagnose_routing.py
@@ -1,10 +1,11 @@
 """
-P0 DIAGNOSE Privacy-First Routing Tests
-========================================
-測試 AIRouter local chain 隔離 + DIAGNOSE timeout 路由
+P0 DIAGNOSE Routing Tests
+==========================
+測試 AIRouter DIAGNOSE 路由 + require_local 隔離行為
 
 建立時間: 2026-04-04 (台北時區)
 建立者: Claude Code (P0 DIAGNOSE Privacy-First)
+2026-04-05 v4.3: Ollama CPU-only 238s 不可用；DIAGNOSE 統一走 NIM (_full_fallback_chain)
 """
 
 import os
@@ -41,7 +42,7 @@ class TestNemotronPerTaskTimeout:
 
 
 class TestLocalFallbackChain:
-    """require_local=True 時只走 local chain，全部失敗 → REJECT，不觸碰雲端"""
+    """require_local=True 時 privacy 過濾生效，cloud provider 不被呼叫；全部失敗 → REJECT"""
 
     @pytest.mark.asyncio
     async def test_require_local_skips_cloud_providers(self):