feat(cli): add jdf eval for RAG retrieval scoring (Recall@K, MRR, nDCG@K)

[mmustafasenoglu]

Summary

Starts on #3 (Benchmarking) by adding the missing measurement piece: a pure retrieval-metrics library plus a jdf eval command that scores a ranked run against relevance judgments (qrels) and writes a JSON report CI can diff over time for regression testing.

What's in here

• tools/jdf-cli/src/lib/retrieval-metrics.ts — pure functions, no I/O: recallAtK, reciprocalRank (mean = MRR), ndcgAtK (exponential gains, log2 discount), plus evaluateRun which aggregates per-query means and skips queries with no judged-relevant docs.
• tools/jdf-cli/src/commands/eval.ts — jdf eval <qrels.json> <run.json> [--k N] [-o report.json]; prints a one-line summary and writes the full per-query report.
• Tests (retrieval-metrics.test.ts, 15 cases on the built-in node:test runner, zero new dependencies) and sample fixtures under tools/jdf-cli/test/fixtures/.
• test script wired into @uurtech/jdf-cli package.json; help text updated.

$ jdf eval eval/qrels.json eval/run.json --k 10
eval: 2 queries @10 — recall 1.000, MRR 0.750, nDCG 0.815
eval: report written to eval-report-k10.json

Verification

• pnpm --filter @uurtech/jdf-cli test → 15/15 pass
• pnpm --filter @uurtech/jdf-cli typecheck → clean
• pnpm --filter @uurtech/jdf-cli build → success
• End-to-end run on the checked-in fixtures produces the expected numbers (hand-verified).

Notes

• CLI-only on purpose, following the existing chunk/embed precedent: this is RAG pipeline tooling, not a format change, so the three-surface rule does not apply.
• Deliberately leaves the eval-set collection, chunker comparison, and CI gating from Benchmarking #3 as follow-ups; happy to take those on next if this direction looks right.

Related to #3