feat(travel): 接入远端 AI 推荐服务，支持 GLM real 模式与本地静默回退

- createPlan 在配置 VITE_TRAVEL_ASSISTANT_API_BASE_URL 时调用 FastAPI 推荐服务
  做无坐标选点，路线/交通/时长仍由本机核算；远端不可用即静默回退本地规划
- 放开 loadPlan 对远端行程的回读限制（原 AI 冻结期禁用），行程页跳转不再丢行程
- planner/assistant 文案按是否启用远端如实切换，不再谎称「本期不调用 AI」
- 新增远端接通/回退/候选载荷单测；server/README 更新联调与 real 模式说明
- .env.development 默认指向已部署推荐服务；server/.env.example 补 GLM 示例

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

server/.env.example

@@ -2,6 +2,9 @@
 LLM_MODE=mock
 
 # Configure these on the server only when LLM_MODE=real.
+# Any OpenAI-compatible endpoint works, e.g. Zhipu GLM:
+#   OPENAI_BASE_URL=https://open.bigmodel.cn/api/paas/v4
+#   OPENAI_MODEL=glm-4-flash   # free tier; use glm-4-plus etc. if you have credit
 OPENAI_API_KEY=
 OPENAI_BASE_URL=https://api.openai.com/v1
 OPENAI_MODEL=

server/README.md

@@ -2,7 +2,7 @@
 
 This FastAPI service is the server-only AI boundary for the Guangming travel POC. It defaults to deterministic `mock` mode. Model credentials belong only in `server/.env` or the deployment environment and must never be added to the mini-program build.
 
-This service is reserved for a later AI integration phase. The current mini-program `createPlan` entry point is hard-wired to the local planner and does not call this service.
+The mini-program calls this service when `VITE_TRAVEL_ASSISTANT_API_BASE_URL` is set (`.env.development` points it at `http://localhost:8000`). If that variable is empty, or the service is unreachable or returns an error, the client silently falls back to its deterministic local planner — so the demo always works. The model only ever recommends POIs from a coordinate-free client whitelist; route geometry, transfer estimates, and final itinerary time remain client responsibilities. The itinerary page labels the result `真实 AI 推荐` (real), `远端演示规划` (mock), or `本地 POC 规划` (local fallback).
 
 ## Contract
 
@@ -60,3 +60,16 @@ Run tests from the repository root:
 ```bash
 corepack pnpm server:test
 ```
+
+## Real model mode
+
+Mock mode is deterministic and needs no credentials. To use a real model, set `LLM_MODE=real` in `server/.env` (gitignored) with any OpenAI-compatible endpoint — for example Zhipu GLM:
+
+```
+LLM_MODE=real
+OPENAI_BASE_URL=https://open.bigmodel.cn/api/paas/v4
+OPENAI_MODEL=glm-4-flash
+OPENAI_API_KEY=<your key>
+```
+
+`GET /api/v1/health` then reports `mode=real, configured=true, ready=true`. With `pnpm server:dev` running, generate a plan from the mini-program (H5 dev server on `:5173` is already in `CORS_ORIGINS`; for `mp-weixin` enable “不校验合法域名” in the devtools). To force the local planner again, unset `VITE_TRAVEL_ASSISTANT_API_BASE_URL` or stop the server.