[FreakyAdy]

Closes #8

Summary

Adds CohortMatrixAuditor to analyze benchmark outcomes across multiple agents and repeated trials. The auditor computes per-agent pass rates, cohort mean and population standard deviation, per-agent z-scores, and a finite-sample bimodality coefficient. It flags universal pass/fail behavior, high dispersion, z-score outliers, bimodal pass-rate distributions, and polarized cohorts, while marking cohorts below the configured minimum as insufficient for anomaly flags.

The result is a structured CohortTaskAnalysis, with an issue-Markdown exporter that includes the per-agent outcome table and an explicit caution that statistical signals are screening evidence, not proof of a defective test. The new classes are exported from evalguard.testaudit.

Tests added

• Universal pass and universal fail cohorts
• A low-performing outlier agent
• Polarized high/low cohort and issue export
• Uniform intermediate outcomes that should not be flagged
• Insufficient cohort size
• Multiple-task matrix handling
• Malformed outcomes and invalid configuration
• Refusal to export an issue when no anomaly was flagged

Validation

GitHub Actions is running for this branch; results will be updated after they complete. This PR is tied to the existing #8 issue and is scoped to its requested cohort-matrix analysis.