ultrai18nlisted

# ultrai18n — find every human-readable string, and prove nothing was missed For engine evidence and coverage limits, read [engine-evidence](references/engine-evidence.md). The hard part of changing a repository's language is not translating. It is **recall**. An agent's model of "the UI" is components plus HTML plus README, and whatever falls outside that model is invisible — not translated badly, simply never looked at. Prompting harder does not fix a blind spot. The opposite failure costs more. `'done'` can be a rendered label *and* a value persisted to storage and validated by a schema. Translate the wrong one and every existing user's data stops parsing. So the engine does the finding and the accounting, deterministically; you do the judgment it refuses to fake; and a model does nothing but translate strings it is handed. > **The core rules:** > > 1. **The claim is accountability, not omniscience.** ultrai18n does not claim to classify every > string correctly. It claims no byte of a scannable file leaves the pipeline unaccounted for. > Anything unclaimed and human-looking becomes `unclassified`, and `check` fails while one remains. > 2. **The engine writes files. Models never do.** A translator receives `{id, text}` and returns > `{id, text}`. It does not open a source file, and it cannot reformat, drop, or "improve" code. > 3. **A refusal is a result.** Where a decision needs code the engine cannot read — a dual-use > string, a plural baked into a tern