Stress–Diffusion Proteostasis Optimization: An Exact-Budget Adaptive Operator Portfolio for Continuous Black-Box Search

A black-box optimizer must decide how to spend a fixed evaluation budget as the search population changes: whether to refine promising solutions, explore new directions, or recover from stagnation. Stress–Diffusion Proteostasis Optimization (SDPO) addresses this decision as an exact-budget operator portfolio. It combines bounded candidate stress, decision-space graph diffusion, global operator credit, successful-direction memory, differential heavy-tailed trials, and chronic-gated replacement. A single objective gateway accounts for every proposal, while the implemented maps preserve bounded state, box feasibility, and an elitist best-so-far record. The confirmatory study comprises 11,520 integrity-verified runs on 24 noiseless COCO/BBOB functions at dimensions 5, 10, 20, and 40, with exactly 10,000d calls. SDPO beats Random Search throughout, is worse than canonical Differential Evolution (DE) at d=5, does not differ under the selected tests at d=10,20, and is better at d=40; CMA-ES-Restart remains stronger. A matched extension shows that L-SHADE is better at every dimension; RecPM-AOS-DE is better at d=5,10,20, with no significant difference at d=40. One- and two-factor ablations, degradation diagnostics, parameter sensitivity, and a transparent in-house CMA-ES audit support the successful-direction archive but show that degradation, rank stress, and current allocation can harm performance. A fixed controller variant, SDPO-R1, improves on SDPO-Full and RecPM-AOS-DE in untouched d=80 BBOB-LargeScale validation but remains behind L-SHADE. An official bbob-noisy check detects neither an overall algorithm difference nor a graph-diffusion robustness benefit. SDPO is also the costliest tested implementation. The contribution is therefore a reproducible adaptive-portfolio architecture and empirical diagnosis with explicit limits, not a claim of universal superiority.