We propose TIG-Diff, a unified graph-diffusion framework designed to address two major limitations of diffusion-based recommender systems: noise-amplified high-order propagation and ranking inconsistency. TIG-Diff integrates implicit subgraph contrastive denoising with curriculum-driven temporal ranking optimization, achieving a synergistic balance between denoising stability and ranking refinement. The Graph Diffusion Denoising Module (GDDM) leverages multi-step diffusion, structural graph embeddings, and temporal dynamics through cross-modal interaction to suppress spurious co-occurrences and adaptively capture evolving user preferences. Meanwhile, Temporal Collaborative Ranking (TCR) introduces curriculum-based optimization that dynamically shifts from MSE-oriented reconstruction to NDCG-guided ranking learning, ensuring a smooth transition from absolute denoising to relative preference modeling. Extensive experiments on four real-world datasets demonstrate that TIG-Diff consistently outperforms state-of-the-art baselines, achieving up to 20.6% improvement in top-K performance. These results highlight the effectiveness of coupling diffusion-based denoising with temporal ranking alignment, offering a robust and generalizable paradigm for next-generation recommendation.