Evals: expand golden set to fourteen domains-spanning queries, add JSON report output

[Ndevu12]

Summary

Roadmap follow-up on the evaluation harness: the golden dataset more than doubles — six queries to fourteen, spanning seven fields — so domain-agnostic ranking quality is finally measured, not assumed. The harness also gains a JSON output mode for automation.

Dataset expansion

Eight new labeled queries with graded relevance and deliberate cross-domain hard negatives:

• federated learning privacy preservation (ML)
• protein structure prediction with deep learning (biomedicine)
• sleep deprivation effects on memory and cognition (neuroscience)
• misinformation spread on social media (social science)
• quantum error correction codes (physics)
• end-to-end deep learning speech recognition (speech)
• gut microbiome influence on brain and behavior (biomedicine)
• generative adversarial networks image synthesis (computer vision)

Dataset integrity policy holds throughout: every candidate is a real paper, and entries whose DOI cannot be stated with confidence carry no DOI rather than a guessed one — the same fabricated-metadata discipline the pipeline itself enforces. Two high-confidence DOIs were added to keep citation validity comfortably above the CI floor.

Measured baseline (keyword-only, fourteen queries)

Metric	Value	CI floor
mean R@5	0.94	≥ 0.80
mean R@10	1.00	—
mean nDCG@10	0.97	≥ 0.85
mean MRR	1.00	≥ 0.85
citation validity	92%	≥ 0.90

All regression floors hold unchanged; the added domains put real headroom pressure on the ranker (per-query R@5 dips to 0.75 on three queries — exactly the targets the reranker should move).

Harness improvements

• Table output gains a domain column and a mean R@5 figure.
• New --json flag on the module entry point emits the full report (per-query metrics plus means) for automation and CI artifacts, backed by format_report_json.
• Docs and the regression-test baseline text refreshed to the fourteen-query numbers.

Verification

• pipenv run pytest tests/ — 427 passed (dataset well-formedness now asserts ≥14 queries across ≥6 domains; new JSON round-trip test), 1 skipped
• ruff check src tests setups — clean; scripts/check_docs_policy.py — passes
• JSON mode smoke-tested end to end: 14 queries, means match the table above