Same-Size Capability Transfer Reveals a Performance-Localization Tradeoff in Deterministic Function Calling

We study whether task-specific function-calling behavior can be transplanted from a donor back into an untuned same-size base model using fixed modules instead of full retraining. Our setting is intentionally narrow and fully deterministic: within-family Gemma models, single-turn function calling, frozen manifests, exact JSON grading, and a primary metric defined on SchemaShift and NoCall stress slices. The donor substantially outperforms the base on the primary strict metric, from 0.0375 to 0.1781 (delta 0.1406, 95% CI [0.1208, 0.1604]). A sparse same-size intervention selected by a locked discovery pipeline reaches 0.2068 strict success on the discovery seed and 0.1658 mean strict success across three seeds, while a matched dense shortcut reaches 0.2177 mean strict success across the same seeds. Sparse structure is nevertheless non-random: a locked one-feature subset retains 47.1% of the full sparse gain and beats random one-feature controls, and one-vector steering is clearly weaker at 0.1078 strict success. The resulting picture is therefore not sparse superiority. Instead, we find a real performance–localization tradeoff: same-size transfer is real in this setting, dense parameter-matched modules recover more raw task performance, and sparse modules provide partial localization signal. We present this deterministic function-calling setup as a concrete testbed for studying constructive capability transfer under tight claim discipline.