16. One data pack per language, and a collision is an error

Context

Entry-point discovery is keyed by language name. Adding a lexicon to a second distribution would have meant two packages registering en, and the loader took whichever loaded first and skipped the rest — silently, so the answers would have depended on installation order.

Decision

Two packs claiming one language raise DuplicatePack, naming both. The WordNet noun list therefore goes into the existing denckring-en-data rather than a second distribution, with each dataset carrying its own licence file: CMUdict under BSD-2-Clause, WordNet under Princeton's licence, code under MIT.

Consequences

The silent failure is gone, which matters more than the packaging convenience it cost: a mistake now announces itself at import rather than changing results invisibly.

The limitation is real and worth naming. A language can have exactly one data pack, so a third party cannot add morphology to English without replacing the pronunciation and lexicon data too. Composition — several packages contributing capabilities to one language — is the fix, and it is future work rather than something to improvise while adding a noun list.

ADR 0013's reasoning still holds for languages: French data will be its own distribution. What has changed is that within a language, data accumulates in one package.