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_providers_test.go

@@ -216,3 +216,107 @@ func TestDebugProviderIsDeterministicAndOffline(t *testing.T) {
 		t.Fatal("debug provider must answer even without history")
 	}
 }
+
+// A reasoning model answers in two parts. When the token ceiling is reached
+// before it stops thinking, the reply field is empty and the whole budget shows
+// up as completion tokens — exactly what the admin log showed as "成功" with a
+// blank reply.
+func TestReasoningModelExhaustingItsBudgetIsRecognised(t *testing.T) {
+	server, _, _ := captureAIRequest(t, `{"id":"call-1","choices":[{"message":{"content":"","reasoning_content":"先想想怎么回答"},"finish_reason":"length"}],"usage":{"prompt_tokens":240,"completion_tokens":256}}`, 200)
+	result, err := chatOnce(t, "openai", server.URL)
+	if err != nil {
+		t.Fatalf("空正文本身不是传输错误: %v", err)
+	}
+	if result.Text != "" || result.Reasoning != "先想想怎么回答" {
+		t.Fatalf("result = %+v", result)
+	}
+	if !result.truncated() {
+		t.Fatal("必须能识别出被截断")
+	}
+
+	// The ordinary case must not be mistaken for it.
+	plain, _, _ := captureAIRequest(t, `{"id":"call-2","choices":[{"message":{"content":"在的，刚吃过"},"finish_reason":"stop"}],"usage":{"prompt_tokens":240,"completion_tokens":62}}`, 200)
+	answered, err := chatOnce(t, "openai", plain.URL)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if answered.Text != "在的，刚吃过" || answered.truncated() {
+		t.Fatalf("result = %+v", answered)
+	}
+
+	// An empty reply with no reasoning and a normal stop is still just empty.
+	blank, _, _ := captureAIRequest(t, `{"id":"call-3","choices":[{"message":{"content":""},"finish_reason":"stop"}],"usage":{"prompt_tokens":10,"completion_tokens":0}}`, 200)
+	empty, err := chatOnce(t, "openai", blank.URL)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if empty.truncated() {
+		t.Fatal("普通空回复不该被当成截断")
+	}
+}
+
+func TestAIRetryMaxTokensStaysWithinBounds(t *testing.T) {
+	for _, test := range []struct{ current, want int }{
+		{256, 1024},  // the configured value that produced blank replies
+		{64, 1024},   // never retry with less than a usable budget
+		{512, 2048},  // four times when that is already sensible
+		{2048, 4096}, // capped, so a misconfiguration cannot bill without limit
+		{8192, 4096},
+	} {
+		if got := aiRetryMaxTokens(test.current); got != test.want {
+			t.Fatalf("aiRetryMaxTokens(%d) = %d, want %d", test.current, got, test.want)
+		}
+	}
+}
+
+// The request the retry sends has to actually carry the raised ceiling.
+func TestChatRequestMaxTokensOverridesTheModelConfiguration(t *testing.T) {
+	server, body, _ := captureAIRequest(t, `{"id":"c","choices":[{"message":{"content":"好"},"finish_reason":"stop"}],"usage":{}}`, 200)
+	provider, err := newAIProvider(testModel("openai", server.URL), false)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err = provider.Chat(context.Background(), aiChatRequest{
+		Messages:  []aiChatMessage{{Role: "user", Text: "在吗"}},
+		MaxTokens: aiRetryMaxTokens(128),
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if got := (*body)["max_tokens"]; got != float64(1024) {
+		t.Fatalf("max_tokens = %v, want 1024", got)
+	}
+}
+
+// A reply that opens with a number used to arrive with the number missing: the
+// list-marker trimming removed the digits themselves.
+func TestSanitizeAIReplyKeepsLeadingNumbers(t *testing.T) {
+	for _, test := range []struct{ raw, want string }{
+		{"30岁啦，你呢", "30岁啦，你呢"},
+		{"1. 先自我介绍", "先自我介绍"},
+		{"2、我在深圳", "我在深圳"},
+		{"- 喜欢跑步", "喜欢跑步"},
+		{"• 也喜欢音乐", "也喜欢音乐"},
+		{"**在的**", "在的"},
+		{"18612725546 是我的号码", "18612725546 是我的号码"},
+		{"3 个月前搬来的", "3 个月前搬来的"},
+	} {
+		if got := sanitizeAIReply(test.raw); got != test.want {
+			t.Fatalf("sanitizeAIReply(%q) = %q, want %q", test.raw, got, test.want)
+		}
+	}
+}
+
+// A cut-off answer is worth one more attempt too, not just an empty one.
+func TestTruncatedAnswerIsRetriedAndKeptIfItSurvives(t *testing.T) {
+	cut, _, _ := captureAIRequest(t, `{"id":"c","choices":[{"message":{"content":"30岁啦，"},"finish_reason":"length"}],"usage":{"prompt_tokens":240,"completion_tokens":1024}}`, 200)
+	result, err := chatOnce(t, "openai", cut.URL)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.truncated() {
+		t.Fatal("被截断的正文也要触发重试")
+	}
+	if sanitizeAIReply(result.Text) != "30岁啦，" {
+		t.Fatalf("text = %q", result.Text)
+	}
+}