feat(pipeline): PromptOptimizer 集成到生成管线，新增优化 API

- PipelineInput 新增 Tags/UserNote 字段
- 管线重构为: START → PromptOptimizer → AssetGenerator → QualitySupervisor → FormatAdapter
- promptOptimizerNode: 有标签调用 PromptAgent，无标签补技术参数，合并风格与重试信息
- PromptBuilder 移除，提示词构建逻辑并入 PromptOptimizer
- 新增 POST /api/v1/prompt/optimize 接口
- main.go 注入 LLM/ImageGen 配置，注册 prompt 路由
- inference.go 新增 InitImageGenConfig 注入

backend/internal/service/nodes.go

@@ -8,13 +8,53 @@ import (
 	"github.com/cloudwego/eino/compose"
 )
 
-// PromptBuilder 节点：接收输入，输出三段式提示词
-var promptBuilderNode = compose.InvokableLambda(func(ctx context.Context, in PipelineInput) (string, error) {
-	return buildPrompt(in), nil
+// promptOptimizerNode 节点：调用 PromptAgent 生成规范提示词，合并风格与重试信息。
+// 输入 PipelineInput，输出最终提示词字符串（直接供 AssetGenerator 消费）。
+var promptOptimizerNode = compose.InvokableLambda(func(ctx context.Context, in PipelineInput) (string, error) {
+	// 合并风格描述，注入原始 Prompt 中
+	styleDesc := buildStyleDescription(in.ProjectStyle, in.TaskStyle)
+	if styleDesc != "" {
+		if in.Prompt != "" {
+			in.Prompt = in.Prompt + "。" + styleDesc
+		} else {
+			in.Prompt = styleDesc
+		}
+	}
+
+	// 注入重试原因
+	if in.RejectReason != "" {
+		if in.Prompt != "" {
+			in.Prompt = in.Prompt + "。注意修正以下问题：" + in.RejectReason
+		} else {
+			in.Prompt = "修正以下问题：" + in.RejectReason
+		}
+	}
+
+	if len(in.Tags) == 0 && in.Prompt == "" {
+		return "", fmt.Errorf("pipeline: Prompt and Tags are both empty")
+	}
+
+	// 有标签时调用 PromptAgent 优化提示词
+	if len(in.Tags) > 0 {
+		agentIn := PromptAgentInput{
+			Tags:      in.Tags,
+			AssetType: in.AssetType,
+			Prompt:    in.Prompt,
+			UserNote:  in.UserNote,
+		}
+		output, err := RunPromptAgent(ctx, agentIn)
+		if err != nil {
+			return "", fmt.Errorf("prompt agent: %w", err)
+		}
+		return output.Prompt, nil
+	}
+
+	// 无标签时直接使用原始 Prompt，补上技术参数段
+	return appendTechNotes(in.Prompt, in.AssetType, in.Params), nil
 })
 
-// promptBuilderPreHandler 首次运行时保存输入到 state；重试时注入 RejectReason
-func promptBuilderPreHandler(ctx context.Context, in PipelineInput, state *PipelineState) (PipelineInput, error) {
+// promptOptimizerPreHandler 首次运行时保存输入到 state；重试时注入 RejectReason。
+func promptOptimizerPreHandler(ctx context.Context, in PipelineInput, state *PipelineState) (PipelineInput, error) {
 	if state.RetryCount == 0 {
 		state.Input = in
 	} else if state.RejectReason != "" {
@@ -23,13 +63,13 @@ func promptBuilderPreHandler(ctx context.Context, in PipelineInput, state *Pipel
 	return in, nil
 }
 
-// promptBuilderPostHandler 将提示词写入全局状态
-func promptBuilderPostHandler(ctx context.Context, out string, state *PipelineState) (string, error) {
+// promptOptimizerPostHandler 将最终提示词写入全局状态。
+func promptOptimizerPostHandler(ctx context.Context, out string, state *PipelineState) (string, error) {
 	state.FinalPrompt = out
 	return out, nil
 }
 
-// AssetGenerator 节点：调用 AI 推理 API 出图
+// AssetGenerator 节点：调用 AI 推理 API 出图。
 var assetGeneratorNode = compose.InvokableLambda(func(ctx context.Context, prompt string) ([]GeneratedImage, error) {
 	var params AssetParams
 	_ = compose.ProcessState[*PipelineState](ctx, func(_ context.Context, state *PipelineState) error {
@@ -39,21 +79,18 @@ var assetGeneratorNode = compose.InvokableLambda(func(ctx context.Context, promp
 	return GenerateImages(ctx, prompt, params)
 })
 
-// assetGeneratorPostHandler 将原始图片写入全局状态
+// assetGeneratorPostHandler 将原始图片写入全局状态。
 func assetGeneratorPostHandler(ctx context.Context, out []GeneratedImage, state *PipelineState) ([]GeneratedImage, error) {
 	state.RawImages = out
 	return out, nil
 }
 
-// QualitySupervisor 节点：质检，输出 PipelineInput 供下游节点消费。
-// 将图片存入 state，设置路由目标 NextNode。
+// QualitySupervisor 节点：质检，设置路由目标。
 var qualitySupervisorNode = compose.InvokableLambda(func(ctx context.Context, images []GeneratedImage) (PipelineInput, error) {
 	var input PipelineInput
 	err := compose.ProcessState[*PipelineState](ctx, func(_ context.Context, state *PipelineState) error {
-		// 保存图片到状态
 		state.RawImages = images
 
-		// 质检
 		style := mergeStyle(state.Input.ProjectStyle, state.Input.TaskStyle)
 		pass, reason, checkErr := CheckQuality(ctx, images, style)
 		if checkErr != nil {
@@ -65,14 +102,13 @@ var qualitySupervisorNode = compose.InvokableLambda(func(ctx context.Context, im
 			state.RejectReason = reason
 		}
 
-		// 决定路由
 		if pass {
 			state.NextNode = nodeFormatAdapter
 		} else if state.RetryCount >= 3 {
-			state.NextNode = nodeFormatAdapter // 超过重试次数，降级输出
+			state.NextNode = nodeFormatAdapter
 		} else {
 			state.RetryCount++
-			state.NextNode = nodePromptBuilder // 重生成
+			state.NextNode = nodePromptOptimizer
 		}
 
 		input = state.Input
@@ -84,7 +120,7 @@ var qualitySupervisorNode = compose.InvokableLambda(func(ctx context.Context, im
 	return input, nil
 })
 
-// formatAdapterNode 节点：从 state 读取图片，格式转换，组装输出
+// formatAdapterNode 节点：从 state 读取图片，格式转换，组装输出。
 var formatAdapterNode = compose.InvokableLambda(func(ctx context.Context, input PipelineInput) (PipelineOutput, error) {
 	var images []GeneratedImage
 	_ = compose.ProcessState[*PipelineState](ctx, func(_ context.Context, state *PipelineState) error {
@@ -119,63 +155,42 @@ var formatAdapterNode = compose.InvokableLambda(func(ctx context.Context, input
 	}, nil
 })
 
-// buildPrompt 构建三段式提示词
-func buildPrompt(in PipelineInput) string {
+// buildStyleDescription 将风格键值对转为自然语言描述，供 PromptAgent 注入。
+func buildStyleDescription(projectStyle, taskStyle map[string]string) string {
+	merged := mergeStyle(projectStyle, taskStyle)
+	if len(merged) == 0 {
+		return ""
+	}
 	var parts []string
-
-	// 【主题】
-	parts = append(parts, fmt.Sprintf("【主题】%s", in.Prompt))
-
-	// 【约束】
-	constraints := buildConstraints(in)
-	parts = append(parts, fmt.Sprintf("【约束】%s", constraints))
-
-	// 【内容】
-	content := buildContent(in)
-	parts = append(parts, fmt.Sprintf("【内容】%s", content))
-
-	return strings.Join(parts, "\n")
-}
-
-// buildConstraints 合并风格 + 负面提示词 + 重试原因
-func buildConstraints(in PipelineInput) string {
-	style := mergeStyle(in.ProjectStyle, in.TaskStyle)
-
-	var parts []string
-	for k, v := range style {
+	for k, v := range merged {
 		parts = append(parts, fmt.Sprintf("%s: %s", k, v))
 	}
-
-	if in.RejectReason != "" {
-		parts = append(parts, fmt.Sprintf("上次质检问题：%s", in.RejectReason))
-	}
-
-	if len(parts) == 0 {
-		return "无特殊约束"
-	}
-	return strings.Join(parts, "; ")
+	return "风格约束：" + strings.Join(parts, "；")
 }
 
-// buildContent 构建技术参数段
-func buildContent(in PipelineInput) string {
+// appendTechNotes 在无标签（不走 PromptAgent）时补上技术参数段。
+func appendTechNotes(prompt, assetType string, params AssetParams) string {
 	var parts []string
-	parts = append(parts, fmt.Sprintf("素材类型: %s", in.AssetType))
-	if in.Params.Resolution > 0 {
-		parts = append(parts, fmt.Sprintf("分辨率: %d", in.Params.Resolution))
+	if prompt != "" {
+		parts = append(parts, prompt)
 	}
-	if in.Params.Frames.Directions > 0 {
-		parts = append(parts, fmt.Sprintf("方向数: %d", in.Params.Frames.Directions))
+	parts = append(parts, fmt.Sprintf("素材类型: %s", assetType))
+	if params.Resolution > 0 {
+		parts = append(parts, fmt.Sprintf("分辨率: %d", params.Resolution))
 	}
-	if in.Params.Frames.FramesPerDirection > 0 {
-		parts = append(parts, fmt.Sprintf("每方向帧数: %d", in.Params.Frames.FramesPerDirection))
+	if params.Frames.Directions > 0 {
+		parts = append(parts, fmt.Sprintf("方向数: %d", params.Frames.Directions))
 	}
-	if in.Params.Format != "" {
-		parts = append(parts, fmt.Sprintf("输出格式: %s", in.Params.Format))
+	if params.Frames.FramesPerDirection > 0 {
+		parts = append(parts, fmt.Sprintf("每方向帧数: %d", params.Frames.FramesPerDirection))
 	}
-	return strings.Join(parts, "; ")
+	if params.Format != "" {
+		parts = append(parts, fmt.Sprintf("输出格式: %s", params.Format))
+	}
+	return strings.Join(parts, "；")
 }
 
-// mergeStyle 合并工程风格与任务风格覆盖，任务同名键覆盖工程
+// mergeStyle 合并工程风格与任务风格覆盖，任务同名键覆盖工程。
 func mergeStyle(projectStyle, taskStyle map[string]string) map[string]string {
 	result := make(map[string]string)
 	for k, v := range projectStyle {