feat(h5): inject direct-chat history on agent escalation and defer ambiguous routing.

Phase B appends snapshot context when leaving h5direct sessions so Goose no longer starts blind. Phase C lets the 0.72 fallback return null on gated chat text so the existing LLM router can run, flags default off.

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

.env.example

@@ -150,6 +150,15 @@ H5_PUBLIC_BASE_URL=http://127.0.0.1:5173
 # MEMIND_CHAT_ROUTER_MEMORY_ENABLED=1
 # MEMIND_CHAT_ROUTER_FALLBACK_ROUTE=direct_chat
 
+# ----- Chat task intent layer（docs/architecture/chat-task-intent-layer-review-20260828.md）-----
+# 阶段 B：direct→agent 升级时把 snapshot 历史注入 Goose 输入（默认关，建议先 canary）
+# MEMIND_DIRECT_ESCALATION_CONTEXT_ENABLED=0
+# MEMIND_DIRECT_ESCALATION_CONTEXT_CANARY_USER_IDS=
+# 阶段 C：0.72 兜底在歧义文本上 defer 给 LLM router（默认关；建议先 SHADOW=1 观测）
+# MEMIND_CHAT_ROUTER_CHAT_SESSION_DEFER_ENABLED=0
+# MEMIND_CHAT_ROUTER_CHAT_SESSION_DEFER_CANARY_USER_IDS=
+# MEMIND_CHAT_ROUTER_CHAT_SESSION_DEFER_MIN_TEXT_LENGTH=12
+
 # ----- Token 压缩（本地开发可选，降低 DeepSeek 上下文）-----
 # Agent 会话历史压缩：超过阈值后清空 Goose 会话并保留摘要（默认 180 条 / 12 万字符，偏松）
 # MEMIND_AGENT_SESSION_COMPACT_MESSAGE_COUNT=40