Few-shot in-context learning is usually evaluated by final accuracy, but correct predictions do not guarantee recovery of the intended task boundary. We formulate few-shot ICL as decision-boundary induction and decompose performance into boundary induction and rule execution. Using controlled binary verification tasks with decomposable boundaries, we diagnose which necessary components are behaviorally enforced and, under chain-of-thought, explicitly recognized. Across three tasks and multiple model families, we show that few-shot ICL often induces component-biased verifiers rather than the full target boundary. Stronger models are more robust to this bias, while chain-of-thought reshapes rather than uniformly improves boundary recovery. An explicit-rule intervention further shows that improvements come from both recovering missing boundary components and executing recognized ones more reliably. These results suggest that evaluating few-shot ICL should go beyond aggregate accuracy and directly test boundary recovery. Code is available at https://github.com/limy805/boundary-icl.