Large language models show promise for source-code vulnerability detection, but a reliable verdict also depends on data dependencies, execution paths, and protective conditions that may be reconstructed inconsistently from source text. We propose VulnPrism, an explainable multi-agent framework for function-level detection. A Strategist selects a small set of security-domain experts, which independently formulate vulnerability hypotheses. A Synthesis Agent groups hypotheses that refer to the same code target into candidates without judging their validity. A Critic then validates these candidates using compact, source-linked control-flow, reaching-definition, dominance, and post-dominance witnesses retrieved from a queryable Evidence Storage. On 435 vulnerable--patched pairs from PrimeVul, VulnPrism achieves 28.28% P-C and a Vulnerability Pair Score of 19.08, improving over the strongest external baseline by 1.61 and 2.53 points, respectively, while reducing its false-positive rate by 1.38 points. The results show that selective expert analysis and evidence-supported validation improve paired vulnerability discrimination without placing the complete program graph in the LLM context.