feat(benchmark): add LoCoMo V2 evidence chain and report - #34

Summary

• add a LoCoMo V2 MiniCPM evaluation path with validated dataset mapping and dialog-mode RawMessage ingestion
• keep chunking, indexing, retrieval, fusion, and reranking daemon-owned; persist ingestion, final retrieval, Answerer, Judge, and diagnostic evidence
• add retry-safe checkpoints, preflight validation, a daemon launcher, and explicit execution-error accounting
• add the completed LoCoMo V2 evaluation report and links to result, trace, session-ingest, and manifest artifacts

Evaluation result

• completed: 1,492 / 1,492 questions
• final execution errors: 0
• LLM-judge accuracy: 931 / 1,492 (62.40%)
• answerer: openrouter:deepseek/deepseek-v4-flash-0731
• judge: openrouter:qwen/qwen3.7-flash
• retrieval: daemon-default Top-12, dialog mode

LoCoMo V2 MiniCPM has no QA-level gold evidence IDs, so Recall@K, Hit@K, Precision@K, and MRR remain null rather than being inferred from session or parent metadata.

Validation

• pnpm --filter @melandlabs/benchmark-locomo typecheck
• pnpm --filter @melandlabs/benchmark-locomo test — 24 tests passed
• pnpm exec prettier --check benchmark/locomo/docs/locomo-v2-test-report.md
• source-scoped Biome lint completed with no errors; the smoke CLI retains 12 expected noConsole warnings
• full 1,492-question run completed with result, trace, session-ingest, and manifest artifacts

Scope note

The final artifact used compatible checkpoint recovery and was produced from a dirty worktree. It is a complete end-to-end diagnostic result, not a clean fresh-database --no-resume baseline for leaderboard or version-to-version comparison.