While domain-specific fine-tuning is widely adopted in medical AI, isolating its impact from architectural differences remains challenging. This paper presents a controlled empirical evaluation of MedGemma-4B against its identical general-purpose baseline, Gemma3-4B, across three text-based and two multimodal benchmarks. Results show that medical fine-tuning provides a consistent and statistically significant performance leap (4.6–11.2 pp, p < 0.01), with the most pronounced gains in complex clinical reasoning (MedQA USMLE) and specialized cardiology (+19.0 pp). Beyond raw accuracy, we uncover critical behavioral shifts: MedGemma-4B exhibits superior positional robustness, successfully mitigating the severe "A-position bias" and prompt-induced collapse observed in the baseline model during Chain-of-Thought reasoning. In multimodal tasks, the model achieved a remarkable +21.0 pp advantage in radiology (VQARAD) by correcting affirmative response tendencies, though no gain was observed in pathology. Despite these advancements, we identify systematic overconfidence as a primary barrier to clinical deployment, as calibration remains poor across all domains. Findings suggest that while 4B-scale domain tuning offers meaningful stability for human-assisted applications, post-hoc calibration is essential for autonomous decision support.