feat: add Hy3 EvalForge MCP server

[soso23412]

Summary

• Add Hy3 EvalForge, a local stdio MCP server for evidence-backed AI-output evaluation.
• Provide specification design, case generation, run scoring, and blinded baseline/candidate comparison tools.
• Enforce artifact path containment, prompt-injection boundaries, secret redaction, deterministic hard checks, evidence validation, and model-call budgets.
• Add public synthetic calibration pairs, real Hy3 validation artifacts, and a terminal demo GIF.

Validation

• 35 passed, 1 skipped
• ruff check .
• ruff format --check .
• python -m build --no-isolation
• python -m twine check dist/*
• stdio MCP tool discovery smoke test

Live validation

• Synthetic baseline/candidate comparison completed with Hy3 and returned BLOCKED for one newly introduced critical disclosure.
• Public 30-pair calibration completed with 90.0% agreement with human reference labels.