This article chronicles six decades of research on metaheuristic optimization in quantitative, bibliometric terms, tracking the trajectory from the evolutionary roots of the early 1960s to the enlistment of foundation models and Large Language Models (LLMs) in optimizer design and tuning. A Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)-style protocol governed screening across five databases (Web of Science Core Collection, Scopus, Institute of Electrical and Electronics Engineers (IEEE) Xplore, Association for Computing Machinery (ACM) Digital Library, arXiv): Of 12,904 records screened, the analysis retains 2,847 core documents published between 1962 and 2025. Descriptive bibliometrics, keyword-drift analysis, and co-citation network analysis (41,203 links; Louvain modularity 0.68) resolve the corpus into six schools of thought: Evolutionary computation foundations (24.6% of the network), swarm intelligence (21.9%), theoretical foundations and landscape analysis (14.8%), benchmarking and empirical methodology (11.2%), constraint handling and real-world engineering (15.3%), and Bayesian, surrogate, and LLM-era optimization (12.2%). Read decade by decade, the thematic record moves from early consolidation through a simulated-physics and tabu wave and a swarm and hybridization wave into the critique- and rigor-centered 2010s, before a data-driven, foundation-model inflection after 2020. A comparative computational vignette races four era markers, Simulated Annealing (SA), Univariate Marginal Distribution Algorithm (UMDA), the Cross-Entropy Method (CEM), and Adaptive Differential Evolution (DE) with Optional External Archive (JADE), on a shared ten-problem testbed, and a random-effects meta-analysis of Hedges' g orders the eras JADE above CEM above UMDA above SA, with family-level reversals. A narrative meta-analysis of 120 focal empirical studies identifies adaptive parameter control (71% of top-quartile improvement reports; claimed gains 10–30%), hybrid local search, diversity management, surrogate modeling, and restart mechanisms as the most frequently invoked performance drivers, under the caution that claimed ranges are self-reported and exposed to benchmark-selection and publication bias. Five Research Questions (RQs), covering intellectual evolution, family dominance, co-citation structure, performance drivers, and the LLM frontier, are posed and answered, and open problems of evaluation integrity, replication, and cost-aware optimization are mapped. A scientometric layer records the field's diffusion: Top-five venue concentration falls from 38.2% to 14.7% while Global South participation rises from 4.1% to 22.6%.