Pipeline (plan → retrieve → gate → judge → aggregate → report/context pack)

Planner

• plan_retrieval generates support + counter queries per claim.
• Respects budgets (MAX_TOTAL_K, MAX_CLAIMS) and domain profiles.
• Emits trace nodes for plan/steps.

Retriever

• Interface: Retriever.search(query, filters, k) -> List[Chunk].
• Optional async interface: AsyncRetriever.asearch(...); async runner uses it when available.
• Filters: CanonicalFilters derived from StateSpec (entity_ids, year, source types, doc_type).
• Adapters provided: LangChain, LlamaIndex, Chroma, Qdrant; implement your own by returning Chunk with provenance populated.

Gate

• Hard eligibility: entity/time match, source policy, freshness.
• Quality/noise: length, boilerplate.
• Diversity: per-source, per-domain, per-doc-type caps.
• Reason codes for every rejection; accepted/rejected evidence recorded in trace.

Judge

• Rule-based (RuleBasedJudge), LLM (LLMJudge), NLI (NLIJudge).
• LLM provider is pluggable (LLMProviderBase); budgets and retries via BudgetedProvider + RetryingProvider.
• Caps chunks per claim and text length to control cost/latency.

Aggregate

• Per-claim: support/contradict scores, coverage, reasons, confidence; primary-source contradictions prioritized.
• Overall: weighted roll-up, critical claim handling, warnings.
• Coverage now tracks doc_types when present.

Trace

• TraceBuilder emits nodes for plan, retrieval, gate decisions, judge calls, evidence assessments, claim verdicts, report/context pack.
• TraceGraph.to_dot() for Graphviz; rejected vs accepted visible.

Outputs

• VerdictReport: JSON + Markdown (citations, accepted/rejected evidence, retrieval plan, metadata).
• ContextPack: facts-first payload for guarded generation.