feat: add IR evaluation benchmark (Recall@K/Precision@K/MRR/NDCG)

[superops-team]

Summary

Adds a reproducible information-retrieval (IR) quality benchmark to okf, quantifying search quality with canonical metrics. This closes the gap identified in the okf-vs-OpenKB analysis: okf previously proved search correctness (functional assertions) but could not quantify search quality (Recall/Precision/MRR/NDCG).

What's included

• pkg/eval — pure-Go IR metric functions: RecallAtK, PrecisionAtK, MRR, NDCG (standard library only, no new dependencies)
• Golden query set — pkg/eval/testdata/golden_queries.json: 20 cases over all 7 document formats (7 single-hit, 5 multi-hit, 3 table-content, 2 list-content, 1 title-match, 2 negative)
• Benchmark runner — RunBenchmark loads the golden set, runs query.Search against a real 7-document bundle (fixtures converted via pkg/convert), scores every case, and outputs aggregate + per-case report
• tools/eval.sh — one-command reproducible benchmark entrypoint (consistent with tools/gauntlet.sh)
• openspec — proposal/design/spec/tasks/conformance (20/20 scenarios fully aligned)
• docs — README Evaluation section + releases.md v1.3.0 baseline scores

Baseline scores (K=5, 20 cases, actual run)

Metric	All cases	Positive only
Recall@5	1.0000	1.0000
Precision@5	0.9000	1.0000
MRR	0.9000	1.0000
NDCG@5	1.0000	1.0000

• 18/18 positive queries: correct top-1, all metrics = 1.0
• 2/2 negative queries: zero results
• All-cases Precision/MRR = 0.9 because 2 negative cases (Precision=0, MRR=0) pull down the 20-case average; positive-only is 1.0

Test coverage

• pkg/eval/metrics_test.go: 19 unit tests + 1 property test (500 random cases verifying all metrics ∈ [0,1])
• pkg/eval/benchmark_test.go: end-to-end benchmark (real fixtures, golden set, baseline assertions) + golden set validation + empty-bundle edge case
• Total: 23 tests, all green

Design decisions

• New package pkg/eval, not extending pkg/query — separation of concerns (retrieval vs measurement), no circular dependency
• Metrics operate on []string (doc identifiers), not []*Concept — pure functions, trivially testable, reusable for future retrieval backends (BM25/vector)
• Binary relevance NDCG — minimal implementation; graded relevance would require extending EvalCase.ExpectedDocs to weighted pairs (documented as known limit)
• okf Search is unordered (results in bundle insertion order) — MRR=1.0 reflects correct first result for these fixtures, not a ranking algorithm; benchmark becomes more meaningful when ranked retrieval is added

Known limits (documented in conformance.md)

1. Binary relevance only (no graded relevance)
2. Small fixture set (7 small docs); large-corpus stress is future work
3. No LLM-as-judge metrics (faithfulness/relevancy) — okf has no generation layer; belongs to future RAG integration

Verification

• go build ./... ✅
• go vet ./... ✅
• gofmt -l (tracked files) ✅ clean
• go test ./pkg/eval/ ✅ 23/23 pass
• go test ./pkg/convert/ ./pkg/git/ ./pkg/lint/ ./pkg/mcp/ ./pkg/okf/ ./pkg/parser/ ./pkg/query/ ./pkg/tool/ ✅ all pass
• tools/eval.sh ✅ exits 0, prints full report
• Determinism ✅ two consecutive runs produce identical scores

Note: cmd/okf CLI smoke tests are slow (~55s each, recompiling the binary every run) and were not re-run locally due to time; they are pre-existing tests unaffected by this change (pkg/eval does not import cmd/okf). CI will verify the full suite.