feat: publish ShellRisk-Bench v0.1

[hasandemirkiran]

Summary

• publish the six-source command benchmark with pinned, download-only adapters
• document provenance, licensing, normalization, and the in-distribution split
• release deterministic split checksums, scoring tools, aggregate results, and hashed k-cl-1 predictions
• exclude classifier weights, serving code, product prompts, raw upstream data, and internal research tracks

Verification

• Python unit tests pass
• all six adapters smoke/build checked
• split manifest reproduced exactly: 4,194 test commands (193 risky)
• k-cl-1 predictions reproduce P=0.9468, R=0.9223, F1=0.9344