Shipping in 76 Locales: A Multi-Case Study of LLM-Agent Localization Pipelines for Production Mobile Apps

Large language model (LLM) agents translate software interfaces at scale, yet production evidence on what breaks—and which safeguards catch it—is nearly absent: a 2026 mapping study of localization testing retained 24 primary studies, though none on LLM-driven localization. We report a multi-case study of fourteen shipped mobile apps by one developer (ten iOS, four Android; 40–76 locales; roughly 347,000 units) at five maturity levels. Mining the repositories (catalogs, 4473 commits, process documents, tooling), we code defects into thirteen classes in two waves, the second under a frozen codebook, and map each to its detecting layer. Translation quality is not independently measured; the defect record is a lower bound. Of 468 commit-anchored defects, process and coverage gaps—strings escaping the pipeline—are modal in each wave (26.7% and 27.5%); placeholder errors are near-absent (9 in 347,000 units). LLM judges are the largest operational detector (26.7%), ahead of humans (16.0%) and validators (13.7%), yet 127 defects (27.1%) escaped every layer until our audit. What escapes tracks validator scope, not maturity alone. Layered validation, not prompt-level care, determines what reaches users. Translation takes minutes; the validation stack is the lasting cost. We derive six testable propositions and ten guidelines, and release the codebook, dataset, and scripts.