Safety alignment of small language models (SLMs) typically incurs an alignment tax: benign prompts are refused along with truly harmful ones. The problem is sharper for Korean SLMs, where limited capacity meets a language whose intent signal is blurred by honorifics. We propose KoSCoPe, which combines a Hierarchical Safety Curriculum, a level-conditional adaptive margin (HARM-DPO), and Counterfactual Consistency Regularization, with a reasoning internalization schedule that recovers 1.5 to 3.2× inference speed. Across five Korean SLMs of 1.7B to 4B, KoSCoPe achieves the strongest Safety/Over-refusal trade-off among the baselines under our matched protocol and remains competitive with general-purpose models several times larger. Under paired-bootstrap tests, its Safety gains over the matched preference-optimization baselines survive multiple-comparison correction for all but the closest competitor, Cat-DPO, where the gain is smaller but directionally consistent. Preliminary cross-lingual validation on Japanese, Mandarin Chinese, and English suggests the same design can transfer beyond Korean, with larger gains where surface politeness markers are more salient.