test: add claude plugin eval suites for quality and dev-standards #25

Goal

quality and dev-standards have no claude plugin eval suites, so CI (.github/workflows/plugin-evals.yml, from #18) skips them and a PR changing their skills shows nothing. github-tools and cad-tools have suites.

Scope

• Cases under plugins/quality-tools/evals/<skill>/<case>/ and plugins/dev-standards/evals/<skill>/<case>/, in the layout github-tools and cad-tools use (prompt.md + graders/).
• Side-effect free: dry runs with read-only tools (allowed_tools: [Skill, Read]), no shell, so they run on a CI runner.
• Each case has a free tool_used: Skill check, plus regex graders where the expected behaviour is a literal command or flag, and an llm grader only where the behaviour needs judging.
• Prioritise skills whose failure costs the most: preflight, create-pr, full-pr-review, act-on-pr, make-ci-green-again, and dev-standards' git-workflow.

Lessons from #18

• Headless runs have no AskUserQuestion; a skill that asks must still ask in its reply (see fix(github-tools): merge-ready asks for scope in the reply when AskUserQuestion is unavailable #23).
• Write llm rubrics so the skill's own fallback advice isn't read as a failure (fix(github-tools): merge-ready asks for scope in the reply when AskUserQuestion is unavailable #23, test(github-tools): anonymous-404 grader accepts the skill's own fallback #24), and check a failing judge against the actual reply before changing the skill.
• CI uses Sonnet 5.5 as both agent and judge; verify new cases locally with --model claude-sonnet-5-5 --judge-model claude-sonnet-5-5 at the default 3 runs.
• Cost: about $0.10–0.25 per case per CI run.