feat: nightly LLM-judged A/B (answer accuracy + hallucination)

[v0ropaev]

Adds the answer-quality evidence on top of Tier-3 recall (DESIGN §9): an answerer LLM answers each question from knowbase's grounded context vs a RAG-over-source context, and a judge LLM scores accuracy (against hand-written gold) + hallucination (claims unsupported by that arm's context). Optional, key-gated, NON-gating — it never blocks CI.

What

• kb.llm — LLMProvider protocol + AnthropicProvider (default, claude-opus-4-8) / OpenAIChatProvider, default_llm_provider() via KB_LLM_PROVIDER, has_llm_key(). Mirrors kb.embed.providers; lazy SDK imports behind a new llm extra.
• tier3_llm_judge_test.py — skipif no key. Builds both context arms (knowbase grounded units + provenance vs RAG chunk text), answers + judges all 11 questions, prints knowbase-vs-RAG accuracy + hallucination and PASS/FAIL vs a pre-registered threshold, writes a metrics JSON. Asserts only that the A/B ran — never the win. Self-judging bias accepted (overridable via KB_LLM_JUDGE_MODEL).
• questions.GOLD — hand-written reference answers for the 11 questions (we own the fixtures → no arm bias).
• RagHit.raw_text — so the RAG arm can feed chunk text to the answerer.
• nightly-llm-ab.yml — scheduled + dispatch, pgvector service, --extra llm, secret-gated ANTHROPIC_API_KEY, continue-on-error, uploads the metrics artifact.
• pyproject — llm = [anthropic, openai] extra + anthropic.* mypy ignore.

Safety / CI

Regular ci.yml stays green: the judge test module imports cleanly without anthropic/openai installed (lazy imports) and skips without a key; no new required deps; mypy covers the SDKs via ignore_missing_imports. Verified locally: 52 passed + 1 skipped, ruff + mypy --strict clean, provider/JSON-parsing smoke OK.

The deterministic Tier-3 recall gate remains the hard floor. After merge I'll trigger the nightly once via workflow_dispatch (with the secret set) to confirm it runs and uploads metrics.