Add retrieval benchmark harness

[TheGreenCedar]

Summary

• Adds holdout/local-real benchmark task manifests and expands the task schema/readme.
• Updates the agent A/B benchmark for mandatory sidecar provenance, packet runtime scoring, and cache preparation.
• Adds value scoring and benchmark-contract scripts with tests, plus a small bench compile fix.

Tests

• cargo check -p codestory-bench --benches
• node --check scripts/codestory-agent-ab-benchmark.mjs
• node --check scripts/codestory-agent-value-score.mjs
• node --check scripts/codestory-benchmark-contract.mjs
• node --check scripts/fetch-holdout-repos.mjs
• node --test scripts/tests/codestory-agent-ab-analyzer.test.mjs scripts/tests/codestory-agent-value-score.test.mjs scripts/tests/codestory-benchmark-contract.test.mjs
• cargo fmt --check
• git diff --check

Stack

Base: codex/accuracy-pass-04-cli-retrieval (#14)
Next: codex/accuracy-pass-06-docs-ci