fix(admission): 限制媒体任务准入锁等待并发

媒体请求从认证、任务创建到异步准入登记共用可配置的请求并发令牌，完成登记后立即释放。这样仍会尽早丢弃 Base64 正文，但不会让高并发请求各占一个 PostgreSQL 连接等待 advisory lock。\n\n扩展 Gemini 双角色压力测试，使用生产同档 API/Worker 连接池并记录数据库连接与 advisory lock 峰值。\n\n验证：go test ./... -count=1；go vet ./...；pnpm openapi；迁移安全检查；256 个 256 KiB 异步 Gemini 请求全部通过，连接峰值 45、锁等待峰值 14。

apps/api/internal/httpapi/gemini_compat.go

@@ -164,13 +164,12 @@ func (s *Server) geminiGenerateContent(w http.ResponseWriter, r *http.Request) {
 		writeGeminiTaskError(status, err.Error(), nil, clients.ErrorCode(err))
 		return
 	}
-	// The queued task only needs the materialized assetRef. Release the native
-	// Base64 request body before it waits for distributed admission or an
-	// upstream response.
+	// The queued task only needs the materialized assetRef. Drop the native
+	// Base64 request body now, but retain the lightweight request slot until
+	// distributed admission is durably registered. This prevents a burst from
+	// turning advisory-lock waiters into one PostgreSQL connection per request.
 	native = nil
 	mapping.Body = nil
-	releaseRequestBody()
-	releaseRequestBody = nil
 	runMode, err := s.admittedTaskRunMode(admission, prepared.Body)
 	if err != nil {
 		writeTaskTrafficError(w, err, clients.ProtocolGeminiGenerateContent)
@@ -237,6 +236,8 @@ func (s *Server) geminiGenerateContent(w http.ResponseWriter, r *http.Request) {
 	runCtx, cancelRun := s.requestExecutionContext(r)
 	defer cancelRun()
 	if streamMode {
+		releaseRequestBody()
+		releaseRequestBody = nil
 		s.writeGeminiGenerateContentStream(runCtx, w, r, task, user, mapping.Model, writeGeminiTaskError)
 		return
 	}
@@ -248,6 +249,8 @@ func (s *Server) geminiGenerateContent(w http.ResponseWriter, r *http.Request) {
 		writeGeminiTaskError(statusFromRunError(err), runErrorMessage(err), runErrorDetails(err), runErrorCode(err))
 		return
 	}
+	releaseRequestBody()
+	releaseRequestBody = nil
 	task, err = s.runner.WaitForTaskCompletion(runCtx, task.ID)
 	if err != nil || !requestStillConnected(r) {
 		return