Eval harness: ranking quality on SWE-bench-Lite + LLM-as-judge

[francomano]

Problem

CoreBall's selector is explainable but we lack a public benchmark — hard to prove token savings vs RAG/chunking and to attract research contributors.

Proposal

Add scripts/eval/ harness:

• Fixture tasks over 3–5 real repos (e.g. mini forks with known relevant files per task)
• Metric 1: Recall@budget (are gold files selected within max_tokens?)
• Metric 2: LLM-as-judge (does a small LLM answer the task better with CoreBall pack vs full scan vs RAG chunks?)
• Metric 3: token savings (% reduction)
• CLI: python scripts/eval/run.py --repo fixtures/repo-a --task \"fix auth\" --baseline rag
• Publish baseline table in docs/benchmarks.md + benchmarks/ badge

Benefits

• Trending: eval, benchmark, llm-as-judge, swe-bench — top GitHub AI search terms in 2026.
• Converts the project from “tool” to “paper-with-code” — invites PRs and citations.
• Unblocks issue Evaluate ranking quality on fixture tasks #7 and Add benchmark suite against real-world repositories #15.

Acceptance criteria

• 3 fixtures with gold file lists (fixtures/*/gold.json)
• scripts/eval/run.py prints markdown table (recall, token saving, cost)
• Optional OpenAI/Anthropic judge (skips if no key, uses heuristics otherwise)
• docs/benchmarks.md updated with results + methodology

Research-good-first-issue, no runtime deps (judge is optional).