fix(openai): 非 OAuth 账号 Chat Completions 也提取 system 消息并注入 instructions

GPT-5/Codex 等模型要求请求中必须包含 instructions 字段。
之前 extractSystemMessagesFromInput + applyInstructions 仅在
OAuth 路径执行，API Key 账号走 Chat Completions 时 system 消息
保留在 input 中但缺少 instructions，导致上游返回 400。

同时在 ChatCompletionsRequest 和 ResponsesRequest 中增加
Instructions 字段，支持客户端直接传入 instructions 透传到上游。

backend/internal/pkg/apicompat/chatcompletions_to_responses.go

@@ -27,8 +27,9 @@ func ChatCompletionsToResponses(req *ChatCompletionsRequest) (*ResponsesRequest,
 	}
 
 	out := &ResponsesRequest{
-		Model:       req.Model,
-		Input:       inputJSON,
+		Model:        req.Model,
+		Instructions: req.Instructions,
+		Input:        inputJSON,
 		Temperature: req.Temperature,
 		TopP:        req.TopP,
 		Stream:      true, // upstream always streams

backend/internal/pkg/apicompat/types.go

@@ -152,6 +152,7 @@ type AnthropicDelta struct {
 // ResponsesRequest is the request body for POST /v1/responses.
 type ResponsesRequest struct {
 	Model           string              `json:"model"`
+	Instructions    string              `json:"instructions,omitempty"`
 	Input           json.RawMessage     `json:"input"` // string or []ResponsesInputItem
 	MaxOutputTokens *int                `json:"max_output_tokens,omitempty"`
 	Temperature     *float64            `json:"temperature,omitempty"`
@@ -337,6 +338,7 @@ type ResponsesStreamEvent struct {
 type ChatCompletionsRequest struct {
 	Model               string             `json:"model"`
 	Messages            []ChatMessage      `json:"messages"`
+	Instructions        string             `json:"instructions,omitempty"` // OpenAI Responses API compat
 	MaxTokens           *int               `json:"max_tokens,omitempty"`
 	MaxCompletionTokens *int               `json:"max_completion_tokens,omitempty"`
 	Temperature         *float64           `json:"temperature,omitempty"`

backend/internal/service/openai_gateway_chat_completions.go

@@ -82,20 +82,35 @@ func (s *OpenAIGatewayService) ForwardAsChatCompletions(
 		return nil, fmt.Errorf("marshal responses request: %w", err)
 	}
 
-	if account.Type == AccountTypeOAuth {
+	{
 		var reqBody map[string]any
 		if err := json.Unmarshal(responsesBody, &reqBody); err != nil {
 			return nil, fmt.Errorf("unmarshal for codex transform: %w", err)
 		}
-		codexResult := applyCodexOAuthTransform(reqBody, false, false)
-		if codexResult.PromptCacheKey != "" {
-			promptCacheKey = codexResult.PromptCacheKey
-		} else if promptCacheKey != "" {
-			reqBody["prompt_cache_key"] = promptCacheKey
+		modified := false
+		if account.Type == AccountTypeOAuth {
+			codexResult := applyCodexOAuthTransform(reqBody, false, false)
+			modified = codexResult.Modified
+			if codexResult.PromptCacheKey != "" {
+				promptCacheKey = codexResult.PromptCacheKey
+			} else if promptCacheKey != "" {
+				reqBody["prompt_cache_key"] = promptCacheKey
+			}
+		} else {
+			// 非 OAuth 账号也需要提取 system 消息并注入 instructions，
+			// 否则上游 GPT-5/Codex 等模型会报 "Instructions are required"。
+			if extractSystemMessagesFromInput(reqBody) {
+				modified = true
+			}
+			if applyInstructions(reqBody, false) {
+				modified = true
+			}
 		}
-		responsesBody, err = json.Marshal(reqBody)
-		if err != nil {
-			return nil, fmt.Errorf("remarshal after codex transform: %w", err)
+		if modified {
+			responsesBody, err = json.Marshal(reqBody)
+			if err != nil {
+				return nil, fmt.Errorf("remarshal after codex transform: %w", err)
+			}
 		}
 	}