Description

Adds a per-language leakage dashboard over one or more benchmark runs so multilingual leakage trends, residual-PHI categories, and the worst labels are visible without exposing source text or report metadata. The dashboard produces both a self-contained sortable HTML report and deterministic companion JSON suitable for CI threshold checks.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature which changes existing behavior)
• Documentation update
• Code refactoring
• Performance improvement
• Test addition/improvement

Changes Made

• Aggregate leakage counts and trends for every wired PII language from benchmark reports or OM-403 heatmaps.
• Render one privacy-safe panel per language plus a sortable summary table, residual category breakdowns, and deterministic worst-label rankings.
• Emit stable JSON with per-language evidence, optional leakage thresholds, combined gate status, and violations.

Testing

• Added tests that prove the feature works.
• New and existing unit tests pass locally.
• Tested synthetic multilingual reports, direct heatmaps, single-language metrics, threshold failures, invalid inputs, and PHI exclusion.

Commands run:

• make format
• make lint
• make format-check
• make type-check
• .venv/bin/python -m pytest tests/ -q — 5,205 passed, 46 skipped

Documentation

• Added Google-style docstrings to the new public functions and classes.
• Updated the changelog (not required for this scoped feature PR).

Code Quality

• Ran make format, make lint, and make format-check.
• Performed a self-review of the complete diff.
• Added no new warnings or dependencies.

Dependencies

• No new dependencies.

Checklist

• Read the contributing and repository guidelines.
• Commit has a clear, descriptive message.
• Branch is based on current origin/master.

Related Issues

Closes #932

Depends on #650, completed by #680.

Screenshots/Examples

The generated HTML is self-contained and includes a sortable language summary followed by detailed language panels. Tests verify that all wired languages are present and that source text, model names, fixture identifiers, and metadata cannot enter either artifact.