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.go

@@ -36,12 +36,21 @@ type aiChatRequest struct {
 
 type aiChatResult struct {
 	Text             string
+	Reasoning        string
 	PromptTokens     int
 	CompletionTokens int
 	FinishReason     string
 	RequestID        string
 }
 
+// The model ran out of budget mid-thought. For a reasoning model that usually
+// means the whole ceiling went on the chain of thought and Text is empty; for
+// any model it can also mean the answer itself was cut off. Both deserve one
+// retry with room to finish rather than a blank or half-finished message.
+func (r aiChatResult) truncated() bool {
+	return r.FinishReason == "length" || (r.Text == "" && r.Reasoning != "")
+}
+
 type aiProvider interface {
 	Chat(ctx context.Context, req aiChatRequest) (aiChatResult, error)
 }
@@ -160,6 +169,20 @@ func aiRequestMaxTokens(req aiChatRequest, model aiModel) int {
 	return 256
 }
 
+// One retry, with enough room for the thinking plus a short answer, and a cap
+// so a misconfigured model cannot bill an unbounded completion.
+func aiRetryMaxTokens(current int) int {
+	const floor, ceiling = 1024, 4096
+	raised := current * 4
+	if raised < floor {
+		raised = floor
+	}
+	if raised > ceiling {
+		raised = ceiling
+	}
+	return raised
+}
+
 func aiPostJSON(ctx context.Context, client *http.Client, endpoint string, headers map[string]string, payload any) ([]byte, error) {
 	body, err := json.Marshal(payload)
 	if err != nil {
@@ -234,6 +257,11 @@ func (p openAIProvider) Chat(ctx context.Context, req aiChatRequest) (aiChatResu
 		Choices []struct {
 			Message struct {
 				Content string `json:"content"`
+				// Reasoning models put their chain of thought here and the answer
+				// in Content. Spending the whole budget on the former leaves the
+				// latter empty, which is indistinguishable from a broken model
+				// unless this field is read.
+				ReasoningContent string `json:"reasoning_content"`
 			} `json:"message"`
 			FinishReason string `json:"finish_reason"`
 		} `json:"choices"`
@@ -256,6 +284,7 @@ func (p openAIProvider) Chat(ctx context.Context, req aiChatRequest) (aiChatResu
 	}
 	return aiChatResult{
 		Text:             strings.TrimSpace(payload.Choices[0].Message.Content),
+		Reasoning:        strings.TrimSpace(payload.Choices[0].Message.ReasoningContent),
 		PromptTokens:     payload.Usage.PromptTokens,
 		CompletionTokens: payload.Usage.CompletionTokens,
 		FinishReason:     payload.Choices[0].FinishReason,