feat(h5): add LLM intent router admin controls and shadow verification.

Expose shadow/canary router policy in ops admin, add FAQ rule fast-path,
and tighten router defaults (1200ms timeout, 0.65 confidence).
Includes verify-h5-llm-router-shadow for production canary rollout.

Co-authored-by: Cursor <cursoragent@cursor.com>

.env.example

@@ -141,6 +141,33 @@ H5_PUBLIC_BASE_URL=http://127.0.0.1:5173
 # 1/on：gateway 优先读 classification.decision.route / session_hint
 # MEMIND_ROUTER_NORMALIZED_DECISION=0
 
+# H5 Chat LLM intent router（本地开发建议见下方「本地 LLM Router」块）
+# 0：仅规则路由；规则未命中 fallback Agent
+# 1 + SHADOW=1：调用轻量 LLM 观测分歧，行为仍 fallback Agent
+# 1 + CANARY_USER_IDS：仅名单内用户在规则未命中时走 LLM 路由；留空=全员
+# MEMIND_CHAT_LLM_ROUTER_ENABLED=0
+# MEMIND_CHAT_LLM_ROUTER_SHADOW=0
+# MEMIND_CHAT_ROUTER_CANARY_USER_IDS=
+# MEMIND_CHAT_ROUTER_MODEL_PROVIDER_KEY_ID=
+# MEMIND_CHAT_ROUTER_MODEL=
+# MEMIND_CHAT_ROUTER_MIN_CONFIDENCE=0.65
+# MEMIND_CHAT_ROUTER_TIMEOUT_MS=1200
+# MEMIND_CHAT_ROUTER_MEMORY_ENABLED=1
+# MEMIND_CHAT_ROUTER_FALLBACK_ROUTE=agent_orchestration
+
+# ----- 本地 LLM Router（pnpm dev，全员 Active，无 Shadow）-----
+# Router 模型建议（分类 JSON，短 prompt，要快不要大）：
+#   本机 Ollama：qwen2.5:3b / qwen2.5:1.5b / llama3.2:3b（见 H5_LOCAL_LLM_*）
+#   云端 relay：gpt-4o-mini / deepseek-chat / qwen-turbo / glm-4-flash
+# Direct Chat 仍用全局选中模型；Router 单独绑小模型即可。
+# MEMIND_CHAT_LLM_ROUTER_ENABLED=1
+# MEMIND_CHAT_LLM_ROUTER_SHADOW=0
+# MEMIND_CHAT_ROUTER_CANARY_USER_IDS=
+# MEMIND_CHAT_ROUTER_MIN_CONFIDENCE=0.65
+# MEMIND_CHAT_ROUTER_TIMEOUT_MS=1200
+# MEMIND_CHAT_ROUTER_MODEL_PROVIDER_KEY_ID=<h5_llm_provider_keys.id>
+# MEMIND_CHAT_ROUTER_MODEL=qwen2.5:3b
+
 # H5 Session stream replay（docs/h5-session-architecture-20260706.md Patch 4c）
 # 默认 0：session SSE 仍纯透传 goosed；设 1 时 Portal 持久化 session 事件并支持 Last-Event-ID 重连补发。
 # MEMIND_SESSION_STREAM_REPLAY=0
@@ -195,7 +222,8 @@ H5_ADMIN_PASSWORD=change-me-admin
 # H5_ADMIN_WORKSPACE_ROOT=/Users/john/PycharmProjects/tkmind
 # LLM Key 数据库加密（可选；默认派生自 TKMIND_SERVER__SECRET_KEY）
 # H5_SETTINGS_ENCRYPTION_KEY=change-me-encryption-key
-# 启动时自动导入 Relay（若不存在同名配置）
+# 启动时自动导入 Relay（若不存在同名配置）；改用手动 Provider 时请设为 1
+# H5_RELAY_BOOTSTRAP_DISABLED=1
 # H5_RELAY_BOOTSTRAP_NAME=Local Goose Relay DeepSeek
 # H5_RELAY_BOOTSTRAP_URL=http://127.0.0.1:18300/relay/buyer/v1/chat/completions
 # H5_RELAY_BOOTSTRAP_API_KEY=your-bearer-token
@@ -305,7 +333,33 @@ VITE_TKMIND_WORKING_DIR=/Users/john/PycharmProjects/tkmind
 # VITE_TKMIND_MODEL=gpt-4o
 
 # 可选：harness recall 的默认查询
-# VITE_TKMIND_MEMORY_QUERY=project architecture conventions current goals decisions risks and recent work
+# Memory V2 Phase A canary (local dev)
+# Full template: docs/memory-v2/phase-a-canary.env.example
+# Readiness check: npm run check:memory-v2-phase-a
+#
+# Recommended local canary users (explicit-memory testers):
+#   32035858-9a20-425b-89da-c118ef0779aa
+#   1c99b83b-0454-474f-a5d2-129d34506a32
+# MEMORY_ENABLED=1
+# MEMORY_CANDIDATE_ENABLED=1
+# MEMORY_CANDIDATE_MODE=canary
+# MEMORY_CANDIDATE_PERSISTENCE_ENABLED=1
+# MEMORY_CANDIDATE_AUTO_ACCEPT_ALL=0
+# MEMORY_AGENT_RESOLVE_ENABLED=1
+# MEMORY_AGENT_INJECTION_MODE=canary
+# MEMORY_AGENT_CANARY_USER_IDS=32035858-9a20-425b-89da-c118ef0779aa,1c99b83b-0454-474f-a5d2-129d34506a32
+# MEMORY_LIFECYCLE_ENABLED=1
+# MEMORY_LIFECYCLE_ROLLOUT_MODE=canary
+# MEMORY_LIFECYCLE_ROLLOUT_USER_IDS=32035858-9a20-425b-89da-c118ef0779aa,1c99b83b-0454-474f-a5d2-129d34506a32
+# MEMORY_PROMOTION_ENABLED=1
+# MEMORY_LIFECYCLE_FORGETTING_ENABLED=0
+
+# Goal Run MVP（默认关闭；canary 用户名单见 docs/architecture/goal-run-data-model-draft.md）
+# GOAL_RUN_ENABLED=0
+# GOAL_RUN_CANARY_USER_IDS=
+# GOAL_RUN_REQUIRE_APPROVAL=0
+# MEMORY_LIFECYCLE_DECAY_ENABLED=0
+
 
 # 历史列表每次展示/加载条数（默认 25，与 desktop 一致）
 # VITE_SESSION_PAGE_SIZE=25