feat(keys): 每个密钥可配独立 AUTO 链，管理员代配

此前 AUTO 链是全局单值（cfg.Auto + Core.AutoChain()），所有用户共用一条链。
管理员无法为某个用户单独指定调度链。

按 per-key 覆盖 + 全局兜底实现：

- config.GWKey 增加 Auto 字段与 HasOwnAuto()。未配置即继承全局链，
  存量部署零改动，新密钥天然继承全局链。
- Core 把 buildAutoChain 的归一化逻辑抽成 chainForRules，全局链、
  生图链、per-key 链共用同一套编译，避免两处漂移。
- Core 增加 keyAutoChains 缓存 + AutoChainFor(key)。请求路径读缓存不
  加锁，与全局链查询一致。缓存整表原子替换，不会看到半成品。
- 请求侧 chat.go 改用 AutoChainFor(reqKey)。冷却仍在 Provider 上按
  model+source 共享：两条链指向同一个 slot 时共用冷却，与今天单链行为
  相同，也避免为 per-key 维度重构冷却而改变现有可观测语义。
- API：GET/PUT/DELETE /api/keys/{key}/auto（admin），GET
  /api/keys/me/auto（任意角色，只能读自己的）。空 PUT 与 DELETE 等价于
  "恢复继承"，无法持久化一条会 503 的空链。写入时回报解析出的槽位数，
  让管理员当场看到模型名写错，而不是等用户下次请求 503。
- 源变更时一并重编译 per-key 链，加源后无需重启即可生效。

判据 18 条，5 个变异全部被抓住：AutoChainFor 忽略 key、清空后不重建
缓存、源变更不重建、空 PUT 落盘成空链、me/auto 误要求 admin。

UI 判据做变异时发现漏放：只查函数定义存在，删掉按钮后仍通过。已改为
断言 keyCanvasHtml 内的调用点。

端到端实测（真实 HTTP + 两个 mock 上游）：admin 与 bob 初始同为 m-fast，
给 bob 配 m-cheap 后两者分流，重启后仍分流，DELETE 后 bob 回到 m-fast。

Co-Authored-By: ModelRouter <noreply@modelrouter.dev>

internal/gateway/keys.go

@@ -7,6 +7,7 @@ import (
 	"strings"
 
 	"llmsproxy/internal/config"
+	"llmsproxy/internal/scheduler"
 )
 
 // handleKeysAPI manages gateway keys: GET /api/keys (admin: all keys),
@@ -17,10 +18,22 @@ func (g *Gateway) handleKeysAPI(w http.ResponseWriter, r *http.Request) {
 	path = strings.Trim(path, "/")
 	role := reqRole(r.Context())
 
+	// /api/keys/me/auto — a user's own chain, readable without admin rights.
+	// Must be matched before the generic {key}/auto case below so it does not
+	// ask admin rights of a plain user asking about their own chain.
+	if path == "me/auto" {
+		g.handleKeyMeAuto(w, r)
+		return
+	}
 	if path == "me" {
 		g.handleKeyMe(w, r)
 		return
 	}
+	// /api/keys/{key}/auto — the per-key AUTO chain editor.
+	if i := strings.LastIndex(path, "/auto"); i > 0 && path[i+len("/auto"):] == "" {
+		g.handleKeyAutoAPI(w, r, path[:i])
+		return
+	}
 	if role != "admin" {
 		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
 		return
@@ -139,6 +152,101 @@ func (g *Gateway) allowedModels(ctx context.Context) []config.ModelScope {
 	return rec.Models
 }
 
+// handleKeyAutoAPI manages one key's own AUTO chain:
+// GET /api/keys/{key}/auto returns it, PUT replaces it, DELETE clears the
+// override so the key falls back to the global chain. Admin only — a user can
+// inspect their own chain through GET /api/keys/me/auto.
+//
+// An empty chain and a cleared chain are deliberately the same state: "inherit
+// the global chain". That keeps the WebUI's "use global" toggle a plain DELETE
+// with no sentinel value to carry through config.yaml.
+func (g *Gateway) handleKeyAutoAPI(w http.ResponseWriter, r *http.Request, key string) {
+	if reqRole(r.Context()) != "admin" {
+		writeError(w, http.StatusForbidden, "forbidden", "admin role required")
+		return
+	}
+	rec, ok := g.core.FindKey(key)
+	if !ok {
+		writeError(w, http.StatusNotFound, "not_found", "key not found")
+		return
+	}
+	switch r.Method {
+	case http.MethodGet:
+		writeJSON(w, http.StatusOK, map[string]interface{}{
+			"key":      key,
+			"auto":     rec.Auto,
+			"inherits": !rec.HasOwnAuto(),
+		})
+	case http.MethodPut, http.MethodPatch:
+		var body struct {
+			Auto []config.ModelScope `json:"auto"`
+		}
+		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
+			writeError(w, http.StatusBadRequest, "invalid_request", "invalid json: "+err.Error())
+			return
+		}
+		if len(body.Auto) == 0 {
+			// Treat an empty PUT as "give up the override" so the endpoint
+			// cannot persist a chain that would 503 with no slots.
+			if err := g.core.SaveKeyAuto(key, nil); err != nil {
+				writeError(w, http.StatusBadRequest, "key_error", err.Error())
+				return
+			}
+			writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true, "inherits": true})
+			return
+		}
+		if err := g.core.SaveKeyAuto(key, body.Auto); err != nil {
+			writeError(w, http.StatusBadRequest, "key_error", err.Error())
+			return
+		}
+		// Report whether the chain actually resolved to usable slots. An
+		// admin who typed a model that no longer exists should learn it here,
+		// not from the user's next 503.
+		chain, _ := g.core.AutoChainFor(key)
+		writeJSON(w, http.StatusOK, map[string]interface{}{
+			"ok": true, "inherits": false, "slots": chainSlots(chain),
+		})
+	case http.MethodDelete:
+		if err := g.core.SaveKeyAuto(key, nil); err != nil {
+			writeError(w, http.StatusBadRequest, "key_error", err.Error())
+			return
+		}
+		writeJSON(w, http.StatusOK, map[string]interface{}{"ok": true, "inherits": true})
+	default:
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "")
+	}
+}
+
+// chainSlots counts the usable slots in a chain, for API feedback after an
+// edit. A chain that compiled to zero slots is reported so the caller can warn
+// instead of letting the next request fail with no_provider.
+func chainSlots(ch *scheduler.Chain) int {
+	if ch == nil {
+		return 0
+	}
+	n := 0
+	for _, tn := range ch.Tiers {
+		n += len(tn.Slots)
+	}
+	return n
+}
+
+// handleKeyMeAuto lets a user read their own AUTO chain without admin rights.
+func (g *Gateway) handleKeyMeAuto(w http.ResponseWriter, r *http.Request) {
+	if r.Method != http.MethodGet {
+		writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "use GET")
+		return
+	}
+	rec, ok := g.core.FindKey(reqKey(r.Context()))
+	if !ok {
+		writeError(w, http.StatusUnauthorized, "invalid_api_key", "key not found")
+		return
+	}
+	writeJSON(w, http.StatusOK, map[string]interface{}{
+		"key": rec.Key, "auto": rec.Auto, "inherits": !rec.HasOwnAuto(),
+	})
+}
+
 // handleAutoAPI manages the AUTO scheduling slots: GET /api/auto returns the
 // current rules; PUT /api/auto replaces them (admin only).
 func (g *Gateway) handleAutoAPI(w http.ResponseWriter, r *http.Request) {