We propose Shape-CFD, a physics-inspired post-retrieval reranking framework that models the relevance redistribution among candidate documents as a conservative convection-diffusion process on a document similarity graph. Unlike pure diffusion approaches, our method introduces a directional advection term derived from the query-document alignment field, enabling asymmetric, query-aware score propagation while preserving total relevance mass through a conservative upwind discretization scheme. We further extend the framework from single-point document embeddings to sentence-level point clouds, replacing cosine similarity with the Chamfer distance to capture fine-grained, sub-document semantic matches. Shape-CFD achieves competitive ranking quality on three BEIR benchmarks, including a 21.5 percent NDCG@10 improvement on NFCorpus, and an 80 percent win rate in blind LLM evaluation on a Chinese statutory retrieval task, while maintaining a reranking latency of 5 ms on commodity hardware.