Single-image flare removal (SIFR) removes lens flare artifacts in degraded images and has become a crucial task in intelligent computing. However, the problem of flare-background entanglement results in a disrupted latent space, where ambiguity in feature attribution emerges and makes it difficult to separate flare degradations from authentic image content. To tackle this challenge, we propose a codebook-based representation approach that transitions from continuous to discrete latent spaces by employing a vector-quantized codebook learned from clear images. This discrete semantic space guides the model to extract flare invariant features by projecting degraded inputs into a prior-informed latent space. Built upon this foundation, we propose a multi-stage architecture that first aligns corrupted features with the codebook priors to reduce semantic inconsistencies, followed by a flare-aware attention module that further refines residual artifacts and enhances fine details. Extensive experiments on both synthetic and real-world datasets demonstrate the effectiveness of our method, yielding notable improvements in PSNR and SSIM compared to existing techniques.