Vision–Language–Action (VLA) models aim to unify perception, language grounding, and robot control, yet recent progress in robotic manipulation remains difficult to compare because systems differ widely in observation modalities, action interfaces, execution regimes, control horizons, and measurement boundaries. The counted analytical corpus comprises 59 unique works: 20 VLA systems in the settings-tuple audit, eight benchmark or evaluation-protocol studies, and 31 additional mechanism or context studies. Survey seeds and general background references are not included in this count. We organize the field through a five-axis design-space taxonomy: (A) perception encoding and spatial grounding; (B) action generation, control regime, and reasoning–control interface; (C) training and adaptation; (D) data provenance and embodiment; and (E) temporal horizon and stabilization. Across these axes, we synthesize five recurring reported limitation patterns—visual brittleness, action degeneration, static supervised fine-tuning limitation, embodiment gaps, and long-horizon breakdown—and consolidate them into four research gaps: reasoning–action decoupling, closed-loop control under contact, spatiotemporal grounding, and data/embodiment mismatch. To support auditable cross-paper interpretation, we introduce a comparability-first reporting schema, a taxonomy coding manual, canonical pseudocode for core VLA mechanisms, and an evidence-tier policy for calibrating claims. We further critique what LIBERO, SimplerEnv, and CALVIN can and cannot establish, showing that current benchmarks remain insufficient for strong deployment-level claims about contact-rich recovery, viewpoint robustness, embodiment transfer, long-horizon stability, and end-to-end latency.