Two derived files, used at build time only by scripts/lib/lemma.js to turn inflected words into dictionary forms for the search index. Neither file is served; what reaches the browser is a ~12 KB map of the corpus's own inflected forms, regenerated on every build.

The lemma index is the half that matters, and it is not the irregular list — it is the validity check. A suffix rule cannot tell string → str from stating → state, because both are "drop -ing, maybe restore an -e". Asking is the result a word? settles it. Rules alone need an arbitrary threshold instead, and measured on this corpus adding one fixed five wrong merges while destroying a right one — use/used/using stopped merging at all. See the header of lemma.js for the full argument.

They are vendored rather than installed because the build must be reproducible: a dependency version bump would silently change the map, and therefore the rankings, with nothing in the diff to show it.

Only needed to move to a newer WordNet. The source is the NLTK corpus distribution, which is the WordNet 3.0 dict/ tree including the exception lists that npm i wordnet-db omits:

git clone --depth 1 --filter=blob:none --sparse -b gh-pages https://github.com/nltk/nltk_data.git nd
cd nd && git show HEAD:packages/corpora/wordnet.zip > wordnet.zip
python3 -c "import zipfile; zipfile.ZipFile('wordnet.zip').extractall('.')"

Then, from wordnet/:

awk '!/^ /{print $1}' index.noun index.verb index.adj index.adv \
  | sort -u | grep -vE '[_0-9]' | grep -E '^[a-z][a-z.-]*$' \
  | gzip -9 > lemmas.txt.gz

exceptions.txt is the four *.exc files concatenated with a leading part-of-speech tag (n, v, a, r), in that order — the order matters, because lemma.js keeps the first mapping for a form and that is Morphy's own precedence.

After regenerating, run npm run check:search-ranking. It pins the merges that must happen, the words that must never be rewritten, and the four traps that a naive rule set reintroduces.

Modern computing vocabulary: namespace, endpoint, microservice, timeout are all absent, and a missing lemma means a correct detachment gets rejected, so namespaces never merges with namespace. That gap is filled by ../project-words.txt, which the build reports candidates for — read the header there before adding any, because roughly one in seven suggestion is a fragment that would break a word which currently works.