Create skill comparison eval extension with HTML report

[ladislas]

Summary

Create a Pi extension that can run evaluation scenarios to compare how the agent/harness/model performs on the same task with and without a specific skill enabled.

The goal is to catch regressions where a skill that is meant to improve behavior actually makes outcomes worse.

Motivation

Skills change agent behavior, but today there is no lightweight way to measure whether a skill improves task performance. A simple eval runner would make skill changes safer by producing comparable results and a reviewable report.

Acceptance Criteria

• Add an extension that can run a task/eval scenario in two modes: with a target skill and without that skill.
• Capture enough output/metadata to compare the two runs meaningfully.
• Generate a simple self-contained HTML report summarizing the findings for human review.
• Keep the first version intentionally small and easy to extend with more eval scenarios later.
• Document how to run the eval extension and where to find the HTML output.