修复 AI 回复为空与开头数字被吞

日志里一批「成功但回复为空」的调用，completion tokens 恰好等于配置的
上限 256：推理模型把预算全花在思维链上，正文没写出来就被截断。

- 解析 OpenAI 兼容响应时读取 reasoning_content，并据此与 finish_reason
  判断「被截断」。
- 被截断时自动加大预算重试一次（4 倍，下限 1024、上限 4096），两次调用
  都记入日志以便核算成本；重试后仍为空才报错，并说明是推理占满预算，
  提示调高 tokens 或改用非推理模型。
- sanitizeAIReply 原本用 TrimLeft 去掉列表符号，连同任何开头的数字一起
  吃掉了——「30岁啦」发出去变成「岁啦」。改为只匹配真正的列表前缀
  （- * • 或 "1. " "2、"），号码、年龄、时长都不再受损。

线上两台已将模型 max_tokens 由 256 调整为 2048。

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

im/backend/internal/app/ai_enqueue_integration_test.go

@@ -4,6 +4,10 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
 	"testing"
 )
 
@@ -168,3 +172,77 @@ func TestDefaultAIModelAcceptsNameOrIDMySQL(t *testing.T) {
 		t.Fatalf("留空时使用标记为默认的一条: id = %d", model.ID)
 	}
 }
+
+// End to end over a real database and a real HTTP model: the first call comes
+// back blank because the reasoning ate the budget, and the reply still has to
+// arrive.
+func TestAIReplyRetriesAfterReasoningExhaustsBudgetMySQL(t *testing.T) {
+	db := isolatedIMDatabase(t)
+	a := &App{db: db, hub: NewHub(), config: Config{Environment: "development", JWTSecret: "isolated-im-regression-secret-only"}}
+	budgets := []int{}
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		var body map[string]any
+		raw, _ := io.ReadAll(io.LimitReader(r.Body, 1<<20))
+		_ = json.Unmarshal(raw, &body)
+		budgets = append(budgets, int(body["max_tokens"].(float64)))
+		w.Header().Set("Content-Type", "application/json")
+		if len(budgets) == 1 {
+			_, _ = w.Write([]byte(`{"id":"a","choices":[{"message":{"content":"","reasoning_content":"想一想"},"finish_reason":"length"}],"usage":{"prompt_tokens":240,"completion_tokens":256}}`))
+			return
+		}
+		_, _ = w.Write([]byte(`{"id":"b","choices":[{"message":{"content":"在的，刚忙完","reasoning_content":"想好了"},"finish_reason":"stop"}],"usage":{"prompt_tokens":240,"completion_tokens":40}}`))
+	}))
+	defer server.Close()
+	for _, statement := range []string{
+		`INSERT INTO system_configs(config_key,config_value,value_type,description) VALUES('ai.enabled','true','boolean',''),('ai.disclose_in_chat','false','boolean','') ON DUPLICATE KEY UPDATE config_value=VALUES(config_value)`,
+		`UPDATE users SET is_test=1,test_batch='regression' WHERE id=2`,
+		`INSERT INTO ai_agents(user_id,model_id,enabled) VALUES(2,1,1)`,
+	} {
+		if _, err := db.Exec(statement); err != nil {
+			t.Fatal(err)
+		}
+	}
+	cipher, err := a.encryptSecret("test-key")
+	if err != nil {
+		t.Fatal(err)
+	}
+	// max_tokens 256 is the setting that produced blank replies in production.
+	if _, err := db.Exec(`INSERT INTO ai_models(id,name,protocol,base_url,model_name,api_key_cipher,temperature,max_tokens,timeout_ms,status,is_default) VALUES(1,'reasoning','openai',?,'reasoning-model',?,0.80,256,8000,1,1)`, server.URL, cipher); err != nil {
+		t.Fatal(err)
+	}
+	var created struct {
+		ID int64 `json:"id"`
+	}
+	if err := json.Unmarshal(imTestCall(t, a.directConversation, 1, "POST", "/api/v1/im/conversations", `{"userId":2}`, 200), &created); err != nil {
+		t.Fatal(err)
+	}
+	if _, _, err := a.persistMessageContext(context.Background(), created.ID, 1, "trigger", 1, map[string]any{"text": "在吗"}); err != nil {
+		t.Fatal(err)
+	}
+	var job aiReplyJob
+	if err := db.QueryRow(`SELECT id,conversation_id,agent_user_id,attempts FROM ai_reply_jobs WHERE conversation_id=?`, created.ID).
+		Scan(&job.ID, &job.ConversationID, &job.AgentUserID, &job.Attempts); err != nil {
+		t.Fatal(err)
+	}
+	if err := a.processAIReplyJob(context.Background(), job); err != nil {
+		t.Fatalf("推理耗尽预算后应当重试成功: %v", err)
+	}
+	if len(budgets) != 2 || budgets[0] != 256 || budgets[1] != 1024 {
+		t.Fatalf("budgets = %v, want [256 1024]", budgets)
+	}
+	var text string
+	if err := db.QueryRow(`SELECT CAST(body AS CHAR) FROM im_messages WHERE conversation_id=? AND sender_id=2 ORDER BY id DESC LIMIT 1`, created.ID).Scan(&text); err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(text, "刚忙完") {
+		t.Fatalf("回复没有落库: %s", text)
+	}
+	// Both attempts are on the record, so the cost of the retry is visible.
+	var calls int
+	if err := db.QueryRow(`SELECT COUNT(*) FROM ai_call_logs WHERE conversation_id=?`, created.ID).Scan(&calls); err != nil {
+		t.Fatal(err)
+	}
+	if calls != 2 {
+		t.Fatalf("calls = %d, want 2", calls)
+	}
+}