In recent years, radar automatic target recognition (RATR) using high-resolution range profile (HRRP) has become a research focus. In real-world, HRRPs frequently exhibit open set and long-tailed characteristics, collectively referred to as realistic HRRPrecognition (RHR). The main challenges in such RHR task include: (1) ensuring that tail classes receive sufficient training for feature representation, and (2) mitigating the overconfidence of deep neural network (DNN) model when encountering unknown classes. To tackle these issues, we propose a synthetic RHR framework (RHRF) for the HRRP-based RATR system. First, we propose an HRRP-conditional diffusion model (HCDM) to generate additional signals for tail-class augmentation. HCDM introduces zero-masked HRRPs as sparse conditional priors to guide the reverse denoising process, thereby achieving a favorable balance between generation stability and diversity. Moreover, the overconfidence issue arises from unconstrained logits magnitudes. To mitigate this, we design two targeted strategies to suppress the excessive growth of logits: a logits-truncation layer (LTL) and a parameter-constraint loss (PCL). LTL directly truncates the maximum value of logits to suppress excessive logits growth, while PCL implicitly limits the logits magnitudes by constraining the DNN parameters. The experiments across diverse scenarios validate the effectiveness and robustness of each targeted module within RHRF.