Researchers have introduced ContextBias, a controlled evaluation framework designed to investigate whether stereotypical associations between professional roles and visual attributes persist when the semantic context is altered. To facilitate this study, the authors developed ContextBench, a comprehensive benchmark consisting of 1,656 semantically controlled prompts covering 92 distinct roles. The study evaluated four state-of-the-art text-to-image models by generating a dataset of 66,240 images across context-free, related, and unrelated semantic conditions. The results indicate that shifting a role into a semantically unrelated context fails to mitigate role-linked biases; instead, cross-role attribute concentration actually increased, evidenced by a pooled bias increase of 0.047. The findings demonstrate that demographic cues, role-specific tools, and characteristic garments remain highly prevalent and robust against semantic prompt reformulation. While scene composition and camera framing exhibited significant context-sensitivity, the core stereotypical attributes remained stable regardless of the prompt's environmental context. This research highlights a form of "stereotypical persistence" that remains undetected by standard context-free evaluations, suggesting that current bias benchmarking methodologies may significantly underestimate the depth of ingrained biases in generative models.