agents(skills): codify maintainer PR evaluation protocol and modernize eval-pr skill

Problem

The current Pull Request evaluation workflow in Deliberate Lab lacks end-to-end codification. While .agents/skills/eval-pr/ provides a mechanical 3-step checkout guide (git fetch + git worktree add), it misses the actual asynchronous maintainer evaluation lifecycle.

In practice, maintainers face several recurring frictions:

1. Implicit Maintainer Workflow: The primary maintainer's end-to-end evaluation procedure (triaging PR scope, gatekeeping manual test steps, belt-and-suspenders local test execution, and behavioral walkthroughs) exists solely in the maintainer's head.
2. Missing Manual Verification Gate: Contributor PRs (which are predominantly new software features) virtually always omit explicit manual testing instructions, creating friction and delays when verifying functionality locally.
3. The AI Agent Communication Gap: Maintainer code reviews are increasingly carried out by or paired with the contributor's AI agent. Review feedback written as vague or conversational comments leads to multi-round thrashing; feedback must be prompt-ready, test-oriented, and unambiguous.
4. Outdated Skill & Missing Invariants: The existing eval-pr skill is purely a mechanical checkout helper. It contains obsolete rules (e.g. prohibiting git branch -D, which breaks for squash-merged remote branches) and lacks automated scaffolding for dependency hydration, local testing, and post-merge retirement (connecting to #1238's automated PR garbage collection).

Proposed Solution

Codify the maintainer evaluation protocol as a formal architecture decision record (.agents/decisions/xxxx-pr-evaluation-lifecycle.md, where xxxx is assigned to the next available number at implementation time) and modernize (or replace/split) the eval-pr skill into a comprehensive maintainer pairing workflow.

1. Architecture Decision Record (.agents/decisions/xxxx-pr-evaluation-lifecycle.md)

Formalize the 6-phase maintainer evaluation protocol into repository canon:

• Phase 1: Triage & Classification:
  • Classify PR scope (Feature vs. Meta/Tooling vs. Bugfix vs. Docs).
  • Trigger / approve pending GitHub Actions CI runs if maintainer approval is required.
• Phase 2: Gatekeeping Verification Pre-requisites:
  • Inspect the PR body for explicit, step-by-step manual testing instructions.
  • If missing (the default case), immediately request them so the contributor (and their AI agent) can draft them in parallel with local test execution.
• Phase 3: Worktree Hydration & Belt-and-Suspenders Local Verification:
  • Spin up isolated sibling worktree: pr-<number>.
  • Hydrate dependencies via npm ci.
  • Compile packages and run lints, unit tests, and integration tests locally to verify test harness integrity independently of GitHub Actions runner quirks.
• Phase 4: Code & Architecture Review (Reviewing for the Contributor's Agent):
  • Inspect diff on GitHub for asynchronous state, multi-party invariants, and test coverage.
  • Formulate review comments specifically as prompt-ready, actionable instructions for the author's AI agent.
• Phase 5: Live Behavioral Walkthrough:
  • Launch local server via ./run_locally.sh inside the pr-<number> worktree.
  • Execute the author's manual testing steps; inspect for regressions and behavioral anomalies.
• Phase 6: Merge & Automated Sandbox Retirement:
  • Squash & merge upstream via GitHub.
  • Rely on workspace-sync (#1238) for automated garbage collection of the evaluation worktree and local tracking branch.

2. Modernize / Refactor the eval-pr Skill (.agents/skills/eval-pr/)

Transform eval-pr from a passive 3-line git checkout guide into an active maintainer assistant:

• Scope & Gate Inspection: Inspect PR metadata (gh pr view) to verify whether manual testing instructions are present; offer copy-pasteable request prompts if absent.
• Automated Worktree Hydration: Automate git worktree add, npm ci, build (npm run build -w utils), and test execution.
• Behavioral Runner Helper: Provide clear invocation commands for ./run_locally.sh within the specific pr-<number> context.
• Align Lifecycle Rules: Update skill documentation to reflect automated terminal-state cleanup via workspace-sync and deprecate the obsolete git branch -d rule.

Related Issues & Dependencies

• #1178 (Automating manual testing step generation via GitHub Actions / AI)
• #1238 (Automated garbage collection of merged/closed pr-<number> worktrees in workspace-sync)
• #1243 (Adoption of .agents/decisions framework)
• #1245 (gh skill helpers for non-truncating PR inspection)

Read the existing .agents/skills/eval-pr/ skill and .agents/decisions framework first, then review related issues #1238, #1243, #1245, and #1178. Trace the current gh pr view, git worktree, npm ci, npm run build -w utils, test, and ./run_locally.sh entry points. Done means the six-phase lifecycle is recorded and the skill covers inspection, hydration, verification, walkthrough, and cleanup.