fix: gateway silently stuck on built-in defaults with zero API keys

A gateway auto-started by `ccmr claude` before any config existed fell back
to DEFAULT_CONFIG with no keys, and stayed that way permanently: the v1.8.0
watcher only observed files that existed at startup, so a models.yaml/.env
created minutes later was never picked up. Because the process is detached,
it survived the terminal and kept answering /health with 200 while every
request 401'd inside Claude Code, with nothing pointing at the real cause.

Four changes, each closing one link in that chain:

- ConfigWatcher polls candidate config/env paths whether or not they exist,
  so files created after startup trigger a reload (replaces watchConfigFiles)
- startup warns loudly when no config file was found or when 0/N models have
  an API key, instead of printing nothing
- `ccmr claude` verifies the reachable gateway can actually serve the launch
  model before handing off, and reports that gateway's config source
- /health exposes config_file and ccmr_home, so "which config is this
  gateway using" is one curl instead of inspecting the process's open files

`ccmr start` is unchanged; auto-start still only fires when the port is free.

Author: 鲁工

README.md

@@ -233,7 +233,13 @@ MiMo Token Plan 的 Base URL 与购买套餐所在集群绑定。默认 `mimo` 
 
 可以自定义供应商、模型变体、别名等。运行 `init` 命令会生成 `providers -> variants` 结构的模板；旧版平铺 `models` 配置仍然兼容。
 
-网关运行中修改 `models.yaml` 或 `.env` 会**自动热重载**（轮询检测，约 1 秒生效），加模型、换 Key 都不用重启。注意：`gateway.host` / `gateway.port` 变更仍需重启才能重新绑定。
+网关运行中修改 `models.yaml` 或 `.env` 会**自动热重载**（轮询检测，约 1 秒生效），加模型、换 Key 都不用重启。**即使网关启动时这些文件还不存在**，之后创建也会被捕获（v1.8.1 起）。注意：`gateway.host` / `gateway.port` 变更仍需重启才能重新绑定。
+
+排查某个网关到底在用哪份配置：
+
+```bash
+curl -s :8080/health | jq '{version, config_file, ccmr_home, default_model}'
+```
 
 #### 故障降级（fallback）
 
@@ -376,7 +382,7 @@ npx claude-code-model-router claude
 | `/v1/messages` | POST | Anthropic Messages API |
 | `/v1/models` | GET | 列出可用模型 |
 | `/usage` | GET | 按模型的用量统计（请求数 / 错误数 / tokens，网关重启后清零） |
-| `/health` | GET | 健康检查（含网关版本号） |
+| `/health` | GET | 健康检查（含网关版本号、`config_file` 配置来源、`ccmr_home`、各模型 Key 状态） |
 
 ## 开发
 
@@ -433,6 +439,17 @@ DeepSeek Anthropic 兼容接口会忽略 `metadata` 字段，但某些 Claude Co
 
 ## 更新日志
 
+### v1.8.1
+
+修复一类隐蔽故障：网关在配置文件存在之前被 `ccmr claude` 自动拉起后，会静默地退回内置默认配置（零 API Key），且此后无论你怎么补配置都不会生效，表现为 Claude Code 里持续报 `401 API key not configured`。
+
+- **热重载改为监视「候选路径」而非「启动时已存在的文件」**：先起网关、后跑 `ccmr init --global`（或手工放置 `models.yaml` / `.env`）现在也能被捕获，约 1 秒内自动生效。此前这种情况下监视列表为空，热重载永不触发
+- **零可用模型时启动告警**：找不到配置文件、或 0 个模型有 API Key 时，启动横幅会明确警告并给出查找路径与修复命令，不再静默
+- **`ccmr claude` 启动前校验模型可用性**：复用已在运行的网关前先检查它能否服务目标模型；不能则直接报错，附上该网关的配置来源与 `pkill` 修复命令，而不是让你在 Claude Code 内部撞上 401
+- **`/health` 新增 `config_file` 与 `ccmr_home` 字段**：一条 `curl` 即可确认某个网关到底在用哪份配置
+
+> `ccmr start` 行为不变。自动拉起只在目标端口探测不到网关时触发。
+
 ### v1.8.0
 
 **易用性**

dist/cli.js

@@ -6,7 +6,24 @@ export interface GatewayHealth {
     status?: string;
     version?: string;
     default_model?: string;
+    config_file?: string | null;
+    ccmr_home?: string;
+    /** model key -> 'available' | 'no_api_key' (absent on gateways older than 1.8.1) */
+    models?: Record<string, string>;
 }
+export type GatewayModelCheck = {
+    ok: true;
+} | {
+    ok: false;
+    reason: 'unknown_model' | 'no_api_key';
+};
+/**
+ * Guards against reusing a reachable-but-useless gateway. A gateway that
+ * started before its config existed answers /health with 200 while holding
+ * zero API keys; without this check the user only finds out via a 401
+ * raised several layers down inside Claude Code.
+ */
+export declare function checkGatewayModel(health: GatewayHealth, modelKey: string): GatewayModelCheck;
 export interface EnsureGatewayResult {
     health: GatewayHealth;
     autoStarted: boolean;