Configuring Spiking Neural Network (SNN) topologies, synaptic parameters, and hardware mappings to minimize energy and latency on event-driven neuromorphic substrates is a decision problem characterized by multimodality, high dimensionality, mixed discrete–continuous variables, and noisy objective evaluations. These properties place the problem beyond the practical reach of classical exact methods and motivate metaheuristic search. This survey synthesizes 95 sources organized into three layers: The metaheuristic methodology itself (evolution-based, swarm-based, physics-based, and human/social-based families, together with hybrid and adaptive schemes), the neuromorphic hardware and SNN training context, and the comparatively sparse intersection literature that applies the former to the latter. Four contributions result: 1) a problem-oriented taxonomy links each algorithmic family to the stages of the neuromorphic design flow, anchored by a formal statement of the architecture optimization problem and a quantitative platform survey, 2) the empirical evidence is consolidated under a uniform assessment protocol—pairwise Wilcoxon signed-rank tests with Holm correction, Friedman ranking with Nemenyi critical-difference analysis, and budget-resolved performance measurement—exposing how strongly reported rankings depend on evaluation budgets, tuning symmetry, and benchmark selection, 3) documented weaknesses in hardware-aware evaluation, noise tolerance, and cross-study reproducibility are traced to their methodological roots, and 4) a five-point research roadmap—surrogate-assisted optimization, Adaptive Operator Selection (AOS), standardized neuromorphic benchmarks, reproducibility safeguards, and algorithm–hardware co-design—formulates concrete actions for each gap. The reviewed evidence indicates that operator-level innovation and methodological discipline, rather than the proliferation of nature-inspired metaphors, offer the most reliable route to gains in energy per inference, synaptic operations, and spike count.