[mforce]

docs-impact.mjs had its own copy of the glossary term rule from glossaryCoverage.ts. Now only one copy remains:

• glossaryCoverage.ts owns termSections (term → section text). glossaryTerms returns its keys.
• docs-impact.mjs imports normalise and termSections from the checkout's glossaryCoverage.ts. In CI, the checkout is the merge commit GitHub creates for the PR. The base revision is still loaded only for its SPEC_ONLY rows, so an older base without the export never matters.
• glossaryCoverage.ts still has no imports, so Node can still load it from any revision.

Closes #1204

Docs-impact: none — CI script refactor, no user-facing strings.

Verification

All checks ran locally at 8ca0685e:

• node --test .github/scripts/docs-impact.test.mjs: 14/14 pass.
• npx vitest run src/routes/helpGlossary.test.ts --maxWorkers=2: 11/11 pass.
• npm run lint and npm run typecheck in web/: clean.
• node .github/scripts/docs-impact.mjs --base origin/main: nothing to flag.
• Parity: glossaryTerms and termSections produce byte-identical output on the current GLOSSARY.md before and after the change (116 terms, none duplicated, so returning the Map's keys changes nothing).
• Mutation: normalise was changed to stop dropping parentheticals. Both checks went red. docs-impact failed 3 of 14: section split, feat(web): prefill the farm code from ?farm= and remember it (#535) #588 drift and wrapped-issue-number. helpGlossary.test.ts failed 3 of 11: every spec term covered, SPEC_ONLY limited to current terms, and every entry names a spec term. With the change restored, both are green.

[mforce]

Review round 1 (Codex, read-only) at 8ca0685e: no findings.

• Import works on CI. glossaryCoverage.ts uses only erasable type syntax, Node 26 runs it with no loader, and older bases only need SPEC_ONLY/NOT_YET.
• No new trust given to the PR. The script already imported head en.ts, helpGlossary.ts and glossaryCoverage.ts. The job has only contents: read, with persist-credentials: false.
• Duplicates. glossaryTerms now returns each name once. Its one caller checks membership and a >100 floor, so nothing relied on duplicates.
• --head. The checker and its rule come from the running checkout, as before, and the comment says so. No guard needed.

Wording nit for the body: CI checks out GitHub's PR merge commit, not the raw head. It still works because the script and the rule come from the same checkout.

VERDICT: 0 product defects, 0 test/doc-only findings.