Sound symbolism, the non-arbitrary mapping between word sounds and meanings, is well known from effects like Bouba–Kiki. But its cross-linguistic scope is difficult to assess because phonological similarity can also arise from shared ancestry, contact, or language-specific inventories. We propose a genealogy-aware stress test for size sound symbolism across 27 languages spanning 13 language families, using 810 phonemically transcribed size adjectives validated with native-speaker audio. We treat cross-family iconicity as a hard generalization task: train on languages grouped by lexical distance to a held-out target, test transfer from increasingly distant languages, and apply adversarial pressure against similarity-bin and source-language cues. Interpretable bag-of-phoneme classifiers predict size above chance, with clear patterns involving /i/ for smallness and /o,a/ for largeness, while consonants add more language-specific cues. Adversarial training prevents the jointly trained model from exploiting genealogy while held-out size prediction stays above chance, though post-hoc probing shows language identity is still recoverable. The residual signal is small but consistent, indicating cross-family size cues.