bench: add fixed SWE-bench Verified N30 cohort

[unohee]

Summary

• add a reproducible, fixed N=30 pure-logic SWE-bench Verified cohort (10 each: pylint, Sphinx, SymPy)
• generate complete official records for the existing sweBench.ts harness
• distinguish low-disk manifest availability from the required sequential amd64 image pull

Validation

• generated all 30 official records from the Hugging Face dataset
• verified 30/30 Docker image manifests
• checked cohort distribution: 10 per repository
• npm run typecheck
• openswarm review --path . — APPROVE

Deliberately not run

• full image pull: local disk had 80GiB free and no SWE image cache; the command is provided but is not claimed as complete
• paid model solves/grading: excluded by the current no-expensive-API-test instruction

[unohee]

🔍 Fresh review of 95772e4 (openswarm pr review --fresh)

✓ Decision: APPROVE
Cohort selection, pagination, validation, and arg parsing are all correct. The data contract (FAIL_TO_PASS as JSON-encoded string) matches sweBench.ts. No production wiring gaps — this is a CLI data generator consumed via JSON file. The three minor observations above are non-blocking.
Suggestions:
- Extract imageFor() into a shared utility if a third consumer appears
- Consider a schema-level test that asserts the output JSON's shape matches sweBench.ts's SweInstance interface