Add SWE-bench Verified adapter

[Bjjj834]

Summary

Adds SWE-bench Verified dataset integration to CodeBench, replacing synthetic sample tasks with real repository-level tasks for benchmark runs. This PR is infrastructure only — it contains no agent automation and no SWE-bench evaluation results yet.

What's included

• Adapter (src/codebench/swebench_adapter.py): loads SWE-bench/SWE-bench_Verified from Hugging Face, converts instances into the existing CodeTask format (no public API changes), and writes prediction files in the official SWE-bench harness format (instance_id, model_name_or_path, model_patch).
• Sampling script (scripts/prepare_swebench_sample.py): writes a small task sample (default 5 tasks; --limit, --difficulty, --output for 30/50-task runs later).
• Prediction-template script (scripts/create_swebench_predictions_template.py): generates a harness-compatible predictions JSONL with empty model_patch stubs.
• Docs (docs/swebench_experiment.md): full workflow — prepare sample → run agents → fill predictions → evaluate with the official harness → map results back into CodeBench metrics.

Leakage prevention

The gold patch, test_patch, FAIL_TO_PASS, and PASS_TO_PASS fields never reach agent-facing output:

• CodeTask.test_file and CodeTask.reference_solution hold placeholders, so a serialized task cannot leak the solution.
• Agents see only the issue description, repo name, base commit, difficulty, and tags (agent_prompt_payload()).
• Gold data is stored solely in a clearly-marked hidden_reference block for internal sanity checks; the generated sample/template files are gitignored.
• A programmatic audit of the generated sample confirmed no gold content in any agent-facing field.

Testing

• 19 new offline adapter tests (fake instances, no network): field mapping, difficulty mapping (<15 min fix→easy … >4 hours→hard, unknown→medium), leakage checks, sampling limits, and official prediction format.
• Full suite passes: 75 passed (56 pre-existing + 19 new).
• End-to-end smoke test against the real Hugging Face dataset verified sampling and template generation.

🤖 Generated with Claude Code