[RooAC]

Summary

containsSimilar used a plain substring check, so deduplicateObservations treated an observation as a duplicate whenever its text appeared anywhere inside another one, even in the middle of a word or number. Distinct facts were silently merged and the shorter one was dropped when they were written through n2n_add_observations / n2n_create_entities:

deduplicateObservations(["API listens on port 80", "API listens on port 8080"])
// before: ["API listens on port 8080"]   ("port 80" is lost)
deduplicateObservations(["Requires Node 1.2", "Requires Node 1.20"])
// before: ["Requires Node 1.20"]
deduplicateObservations(["cat", "Uses a concatenation step"])
// before: ["Uses a concatenation step"]

The containment check now only matches when the contained text does not start or end in the middle of a Latin word or a number. Intended merges still work ("version 2.4" vs "version 2.4.1", "version 2.4.1" vs "version 2.4.1 released"), and text written without spaces (e.g. Chinese, 使用redis vs 使用redis缓存) keeps plain substring matching.

Testing

• Added tests in src/test/similarity.test.ts; the new word-boundary tests fail on main and pass with this change.
• npm run lint, npm test (111 passed), npm run build and npm run pack:dry-run pass locally.
• npm run check stops at npm run audit because of existing advisories in transitive dependencies of @modelcontextprotocol/sdk (body-parser, fast-uri, qs); this is also the case on main and is unrelated to this change.

Checklist

• Tests were added or updated where needed.
• Documentation was updated where behavior changed. (CHANGELOG.md and docs/CHANGELOG_zh.md)
• npm run check passes locally. (see audit note above)