feat(research): rank and deduplicate news sources

[yc2bgr8]

Summary

• add deterministic URL canonicalization, title/content fingerprinting, duplicate and syndicated-source clustering
• rank cluster representatives by explainable source class, freshness, relevance, and availability signals
• enforce Top-K domain diversity and persist policy/version/selection reasons in the Deep Research data bundle
• add a live public-search smoke command plus implementation notes and boundaries

Research integration

ResearchRunner now normalizes research.news results before synthesis. Selected items retain canonical URL, source class, cluster id, and quality score. A sibling research.news:ranking record preserves cluster membership and selected/dropped reasons so duplicate URLs cannot masquerade as independent evidence.

Verification

• bun run typecheck — passed across all workspaces with Bun 1.4.2
• focused tests — 17 passed / 0 failed
• full bun test --isolate — 1221 passed / 5 failed on this Windows host
• latest origin/main baseline — 1207 passed / 6 failed; the branch introduces no new deterministic failures
• existing Windows failures cover symlink permission, Unix-path assertions, and locale-dependent formatting; the flaky ResearchService temp-directory race was independently reproduced on origin/main
• git diff --check — passed

Live search smoke

Command:

bun run eval:retrieval-ranking-live -- "NVIDIA earnings"

Observed on 2026-09-11:

• 49 raw public search results
• 49 independent clusters
• 10 selected sources
• SEC filing ranked first, with Reuters/CNBC/WSJ/Bloomberg results subject to the two-per-domain diversity limit
• exit code 0

Scope

This is an incremental foundation for #39. Redirect fetching, full article-body similarity, Source Inspector UI, and claim-verifier integration remain follow-up work, so this PR intentionally does not close the entire issue.

Refs #39

[helsome]

验证材料足够：Bun 1.4.2、Windows、全仓 baseline 对照、live public-search smoke 的实际命令与结果都给出了；本 PR 也明确只作为 #39 的增量基础，不提前关闭整个 Issue。代码实现方向可合。当前仅因为最新 main 已前进导致分支 dirty，请 rebase/resolve 后跑基础 CI 即可，不需要重做 live smoke。