Watcher: quote arXiv phrases and fix hint ids — 27 of 39 recent issues were noise

[emmanuelgjr]

Plan ID: — (follows the watch backlog triage) Ticket: — Wave: —
Constraints honored:

• C4: the watcher's hints are triage suggestions on an issue, not mappings. Ids were corrected to match the entry titles the rules already named.
• C2: no docs/ change.
• C1: none.

Stacked on #37, because the new test file triggers the CI jobs that are red on main (Drift 11). Merge #37, then retarget to main. This PR touches no generated file, so it won't conflict with #87/#88/#90.

Why

The watcher opened 39 arXiv issues between 2026-08-31 and 2026-09-14. Triage closed 27 of them as out of scope (CBDC settlement, 6G NOMA, MIMO lattice hardness, TPM measured-boot tracing…). All carry triage:out-of-scope with a reason.

Cause: unquoted multi-word terms. In the arXiv API ti:LLM+security is ti:LLM plus a bare security term matched in any field, so the query admitted nearly all of cs.CR.

Live check, 2026-09-14, 30 newest results	AI-related titles
Current query	3 / 30
This PR's query (quoted, title-scoped)	30 / 30

The issue bodies' Suggested OWASP Mapping was wrong too:

• No word boundaries: /rag/ matched "storage" and "average".
• Stale ids, pre-2026 numbering: supply chain → LLM10 (now Improper Output Handling), poisoning → LLM04 (now Supply Chain). For example, [new-research] arXiv:2609.04820 — Cost-Aware Hierarchical Multi-Agent Ransomware Detection and Family Attribution #78 (ransomware detection) was hinted as DSGAI04, LLM04, ASI06, ASI01, ASI02.

Changes

• scripts/watch.js
  • arXiv query with quoted phrases: "prompt injection", "large language model", "AI agent", "retrieval-augmented", plus jailbreak, LLM, agentic, RAG, MCP, still inside cat:cs.CR.
  • ARXIV_HINT_RULES: word-bounded patterns, each naming the entry title it targets.
  • main() guarded by require.main === module, and the query/rules exported for tests. Running the file directly behaves as before.
• scripts/watch.test.mjs (6 tests)
  • Multi-word terms are quoted, plus a test that runs the same check on the old query and expects all four defects.
  • The query stays in cs.CR.
  • No substring matches ("storage", "toolchain").
  • 2026 ids are used.
  • Every hint id exists in data/entries and still carries the title its rule names, so the next renumbering fails CI instead of silently mis-hinting.

Dry run (node scripts/watch.js --dry-run --watcher arxiv): the 20 papers returned are LLM, agent or RAG security topics.

Deliberately NOT changed

• The NVD, OWASP-repo and framework-page watchers. The ATLAS "content change" ([framework-update] MITRE ATLAS — content change detected #83) fired on a homepage ETag, which is a weak signal. Pinning ATLAS releases instead is proposed in MITRE ATLAS: 81 of 117 mapping rows pair an id with another technique's name; 6 ids do not exist (checked against ATLAS 2026.08) #93.
• weekly-watch.yml.

Noted while testing

While testing, I loaded the previous watch.js with require(). It had no require.main guard, so it ran all watchers once from this machine. With no GITHUB_TOKEN set it opened nothing; the issue list confirms none were created. It wrote no state file. This PR adds the guard that would have prevented that.

Verify

$ node --test scripts/watch.test.mjs   6/6
$ npm test                              57/57
$ node scripts/validate.js              0 error(s), 84 warning(s), 312 passed  (unchanged)

🤖 Generated with Claude Code

https://claude.ai/code/session_014SfR2YzLxRH54DAVzDk8gR