This survey addresses the lack of reliable evaluation frameworks for LLM-based financial multi-agent systems (MAS). The authors introduce a four-dimensional taxonomy covering architecture pattern, coordination mechanism, memory architecture, and tool integration, applied to 12 multi-agent systems and two single-agent baselines. They identify five systematic evaluation failures—look-ahead bias, survivorship bias, backtesting overfitting, transaction cost neglect, and regime-shift blindness—that can reverse the sign of reported returns, making cross-system comparisons unreliable. Building on structural patterns that survive these failures, they formulate the Coordination Primacy Hypothesis (CPH): inter-agent coordination protocol is the most consequential design factor for trading decision quality, often more important than model scaling. The CPH is presented as a falsifiable hypothesis supported by tiered evidence (live-market benchmarks, ablation studies, theoretical arguments) but not yet empirically validated due to evaluation infrastructure gaps. To operationalize the CPH, they introduce the Coordination Breakeven Spread (CBS) metric, which determines whether coordination adds genuine value net of transaction costs by comparing coordination-driven signal improvement to the bid-ask spread. The paper concludes with minimum evaluation standards (contamination control, point-in-time universe, rolling-window reporting, net-of-cost returns, regime coverage) as prerequisites for validating the CPH, and identifies three future directions: controlled CPH validation, small language model specialist architectures, and systemic risk from correlated AI trading.