Add reproducible Weighted Mean calibration benchmark

[sylvesterkaczmarek]

Fixes #48.

Summary

Add a reproducible calibration baseline for the Weighted Mean detector across effective token lengths. The benchmark uses deterministic synthetic g-value fixtures, calibrates thresholds at requested false-positive rates, and reports ROC-AUC plus TPR with bootstrap confidence intervals in a versioned JSON result.

The synthetic fixture is explicitly documented as a calibration/reporting baseline rather than a production-performance dataset, so a public text corpus and benign transformations can be added later without changing the metric helpers or result shape.

Testing

• pytest -q — 20 passed
• pyink --target-version py39 --check benchmarks/detector_eval/benchmark.py benchmarks/detector_eval/benchmark_test.py
• benchmark CLI smoke run with two length buckets and JSON output
• git diff --check

[sylvesterkaczmarek]

Refreshed onto current main; the only conflict was README wrapping around the Weighted Mean detector guidance. I retained the current upstream wording and added the benchmark link.

The branch is now 0 behind, mergeable, and GitHub Verified. The detector calibration benchmark tests pass 4/4 on a CPU JAX environment, and git diff --check is clean. Google CI is currently awaiting maintainer approval to run.

Could a maintainer review the refreshed head when convenient?

[sylvesterkaczmarek]

Refreshed onto current main. The branch is now 0 behind, mergeable, and GitHub Verified. The benchmark remains a single focused commit; GitHub CI and the Google workflow are currently at the maintainer-approval gate. Could a maintainer re-review/approve the refreshed head when convenient?