Automated wood surface defect detection is difficult to evaluate reliably because defects are often small, low-contrast, and visually confounded by natural wood texture, while reported performance can vary substantially with benchmark design and domain shift. To address this issue, we conduct a comparative study across three practically relevant settings: a curated seven-class benchmark, a broader in-domain seven-class protocol derived from the same source dataset, and supervised adaptation to a low-resource Vietnamese target domain. We compare lightweight two-stage detectors based on Faster R-CNN with MobileNetV3-FPN against a compact YOLOv8s baseline, while also testing two small-object-oriented YOLO refinements as targeted diagnostic variants rather than as the primary claimed contribution. Across in-domain experiments, the compact YOLOv8s baseline delivers the strongest performance, achieving 84.38% AP50 on the curated benchmark, whereas performance drops to 81.16% AP50 under the broader protocol, indicating that benchmark breadth materially changes the apparent difficulty of the task and the relative strength of competing models. In the target-domain setting, source-initialized fine-tuning improves optimization behavior and can outperform target-only training in a representative single run, but repeated-seed evaluation does not confirm a stable held-out-test advantage under the same adaptation budget. These findings suggest that conclusions drawn from a single curated benchmark may overstate model robustness, and that for wood defect detection, protocol breadth and source-to-target shift should be treated as central evaluation factors rather than secondary experimental details.