feat: lexical citation verifier — flag bogus [N] cites in synthesized answers

[askalf]

Summary

After the final synthesis, deepdive now checks every [N] reference in the answer body against the extracted text of source N. Catches the dominant failure mode of cited-answer tools: a confident sentence pointing at a source that doesn't actually support it.

• Pure-function lexical scoring, no second LLM "judge" pass (would reintroduce the very hallucination class we're catching) and no new runtime deps.
• Multi-cite rule: require all. A sentence with [1][3] is supported only when every cited source clears the recall threshold — a bogus [3] buried in an otherwise-true sentence is still flagged.
• Tokenization: lowercase, drop ~80 stop-words, split on non-alphanumeric and digit/letter boundaries (5h → ["5","h"], 5-hour → ["5","hour"]) so numeric anchors survive paraphrase. Numbers and length-≥3 alphas survive; everything else gets dropped.
• Quiet by default. Clean runs print nothing extra. When something fails: a small ## Citation health footer in the markdown, one warning line per unsupported sentence in --verbose, a verification key in the --json payload.

What's added

• src/verify.ts — verifyCitations + four exported helpers, all pure
• New CLI flags: --strict-cites, --cite-min-recall=<0..1> (default 0.4), --no-verify-cites
• Env vars: DEEPDIVE_STRICT_CITES, DEEPDIVE_CITE_MIN_RECALL, DEEPDIVE_NO_VERIFY_CITES
• New verify.done agent event
• AgentResult.verification: VerificationReport | undefined + usage.{citationsTotal, citationsSupported} for library consumers
• README "Citation verification" section explaining what it catches and what it doesn't (explicitly: not a semantic judge — flags hallucinated names/numbers/dates with high precision, can flag paraphrased-but-truthful sentences below threshold)

What it explicitly is not (v1)

• A semantic / embedding-based judge
• Run inside the deep loop (post-final-synthesis only — feeding unsupported cites back to the critic is a v2 conversation)
• Detection of uncited claims (orthogonal feature)

Test plan

• 39 new unit tests in test/verify.test.mjs covering all five exported helpers + faithful/hallucinated/multi-cite/threshold/strip-sources cases
• 2 new agent-loop integration tests: end-to-end bogus-cite flagging (synth output cites [1] for content not in the source → unsupported.length === 1) and verifyCitations: false skip path
• 4 new CLI/config flag-plumbing tests
• npm run typecheck clean
• npm run build clean
• npm test — 251/251 passing (up from 212)
• --help smoke test shows the three new flags