Add ranked-list retrieval benchmark mode

[lilyjge]

Summary

• add an OUTPUT_MODE=ranked_list path that prompts agents for one deduplicated ranked docid list, persists ranked_docids, enforces the requested depth, and writes TREC run files
• propagate ranked-list options through single, shared, and sharded launchers, including merged TREC output
• add query-set-specific trec_eval metrics and full MS MARCO v1 passage dev setup alongside DL19 and DL20
• document the GPT-5.5 Pyserini REST 2-tool runs and comparison protocol

Recorded results

Query set	AP	nDCG@10	RR@10	R@1K
DL19	0.3780	0.7381	n/a	0.4811
DL20	0.4492	0.7412	n/a	0.5376
MS MARCO passage dev	n/a	n/a	0.3300	0.6725

These runs use openai-codex/gpt-5.5, medium thinking, the Pyserini REST msmarco-v1-passage index, and the pyserini-rest-2tool interface (search and read_document). Raw run and evaluation artifacts remain ignored; the protocol and final values are recorded in docs/msmarco-v1-passage-ranked-list-results.md.

The agent emits roughly 23-27 documents per query while the reference systems retrieve 1,000. AP, nDCG@10, and RR@10 are useful ranking comparisons; R@1K is reported but is not depth-matched and should not be treated as a fair recall comparison.

Verification

• npm run check
• npm test (230 passing)
• bash -n scripts/benchmarks/msmarco_v1_passage/setup.sh scripts/benchmarks/msmarco_v1_passage/generate_query_slices.sh
• npm run prek