v0.6.0 - Long-Form Posts & Eval Depth

"Longer posts, harder test cases, a cleaner toolchain."

Adds long-form post classification, sharpens the evaluation set, and clears the extension's dev-dependency security debt. No breaking changes.

New

• Twitter Article (Notes) support: long-form posts at /i/notes/ URLs are now extracted and classified, not just tweets.

Fixed

• Extracted post text is now whitespace-normalized before hashing, so the same post with different internal spacing no longer misses the cache and gets re-classified.

Changed

• Sharper evaluation set: added fearmongering, divisive/ragebait, and YouTube hype/clickbait boundary cases, plus more YouTube and Reddit examples for platform coverage (@vaishnavidesai09, #92).
• Refreshed the accuracy claim to 96% - the old 98% was measured on a smaller, easier set; the current set is larger with harder boundary cases, and 8B+ models now lead, so llama3.1:8b is the recommended default.

Security

• Cleared all 21 npm audit advisories in the extension dev toolchain (all dev-only and transitive under jest; the shipped extension carries no node_modules). No runtime or user exposure.
• THREAT_MODEL.md gained an engineering-controls table, and the chrome-extension://* CORS note was corrected: it is a no-op literal that fails closed, not an over-permissive wildcard (the extension reaches the server via MV3 host_permissions).

Infrastructure

• The extension Jest suite (81 tests) now runs in CI, so JavaScript regressions are caught. Docker entrypoint hardened (PUID/PGID drop via gosu, SIGTERM forwarding). Dependabot PRs no longer trip the PR-description check.

---

git pull