feat(ml-worker): add BETH Tier 2 sanity rail

[Xore]

Summary

• Add the offline BETH Tier 2 sanity-rail command with the authors' feature encoding and published split fingerprint checks.
• Keep BETH explicitly parallel-corpus-only: AUROC headline, AUPRC alongside, five-or-more seed variance, runtime MD5 metadata, and no disposition or deployment dependency.
• Add focused synthetic tests, documentation, and external-data ignore rules.

Verification

• python3 -m pytest ml-worker/tests/test_detector_benchmark.py ml-worker/tests/test_beth_benchmark.py -q — 38 passed.
• CLI help succeeds without the dataset.
• Empty --data-dir exits 2 with a clear dataset-absent error and no traceback or report file.

The BETH corpus was not downloaded, so real AUROC/AUPRC values and runtime compatibility with the corpus remain to be measured on a prepared external dataset.

Closes #3294

[github-actions]

Dependency Review

✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.

Scanned Files

None