Add `issue-checks` action

where it does:

• semantic duplicate detection:

  • use an embedding model (ex. OpenAI text-embedding-3-small) to convert the new issue into a vector.
  • vector save as cache.
  • compare with embeddings of existing open issues (titles + descriptions).
  • compute cosine similarity (if above $$N$$ threshold, flag as potential duplicate). (custom)

• action:

  • comment on the new issue linking to similar ones. (custom)
  • Llabel it duplicate/needs-info/straight close as duplicate. (custom)

alternatives:

that might be a bit too over-engineered. i'm thinking of trying a prompt-to-search approach from the issue instead, gh issue list --state=open --limit=500 --json=title,url,body --search="X sort:created-asc", where $$X$$ is an AI-generated query.