Add portable local PR review workflow

[vnv-varun]

Summary

This PR implements Nick's draft proposal in #6: make the PR review system portable so the same review shape can run locally before a PR exists.

Today, the strongest review machinery in team-skills assumes a GitHub PR already exists. This change makes that review flow usable before push and before PR creation, so an autonomous /ship run can review its own branch locally, fix validated issues, and only then open a draft PR.

Concretely, the new behavior is:

• stage the review bundle into the target repo
• generate PR context from local git state
• run the multi-agent review flow locally
• write a markdown review artifact to tmp/ship/review-output.md
• let /ship use that artifact as a bounded pre-push gate

Key decisions

• Keep reviewer selection and adjudication shared between GitHub and local modes. Local mode changes delivery only: it writes markdown findings instead of pending review comments.
• Stage the review bundle into tmp/ship/pr-review-plugin and execute the staged copy from the target repo so feature repos do not need to vendor ci/pr-review.
• In Docker mode, run local review from a slim temporary workspace copy that excludes heavy local-only trees like .claude/worktrees; this avoids container hangs from repo-local Claude state while preserving git state and copying review artifacts back.
• In Docker mode, prefer a validated Anthropic API key first, then fall back to host-backed claude.ai credentials only if no valid API key is available.

What changed

• Added ci/pr-review/agents/pr-review-local.md and ci/pr-review/agents/references/pr-review-core-workflow.md.
• Added local review scripts: detect-target-branch.sh, generate-pr-context.sh, and pr-review.sh.
• Added ship helpers: run-local-review.sh and stage-local-review-bundle.sh.
• Updated pr-review-output-contract for local repo-relative citations.
• Updated /ship docs, stop-hook guidance, and PR-creation timing to make the local review gate a bounded pre-push step.
• Updated ci/pr-review/README.md to document the new local entrypoints and durable run artifacts.
• Added wrapper hardening so a valid local review artifact can be treated as success even if Claude exits nonzero after writing it, while still recording the raw exit code in run metadata.
• Added Docker auth hardening so run-local-review.sh validates API key sources, prefers a working repo or shell API key, and falls back to host-backed claude.ai credentials only when needed.

Validation

• bash -n on the new and updated shell scripts.
• End-to-end Docker-mode run from code/agents on branch codex/preview-auto-followup-2440-clean using .ai-dev/docker-compose.yml.
• Verified the run stages the bundle, auto-detects main as the target branch, generates local pr-context, writes tmp/ship/review-output.md, and copies durable artifacts back under tmp/ship/local-review-runs/.
• Verified the Docker path now runs Claude from /tmp/local-review-workspace instead of the mounted /workspace repo, which avoids recursing through the repo's 9.1 GB .claude/worktrees tree during startup.
• Validation artifact from the successful writeback case: code/agents/tmp/ship/local-review-runs/20260317T134310Z-codex-preview-auto-followup-2440-clean-9410/review-output.md.
• Traced the earlier Docker auth failure to a stale shell ANTHROPIC_API_KEY that overrode a different valid key in code/agents/.env.
• Validated the repo-local Anthropic key against https://api.anthropic.com/v1/models.
• Verified direct sandbox Claude execution with the validated repo key via claude -p "reply with OK only" --dangerously-skip-permissions --max-turns 1, which returned OK.
• Verified a full API-key-backed local review pass completed and produced a real markdown review summary under the corrected auth path.

Notes

• Nick's PR proposal: portable code review system #6 was a design proposal, not an implementation. This PR is the implementation of that design plus the Docker and runtime hardening needed to make it work against the real code/agents isolated environment.
• The review findings emitted by the end-to-end validation are findings on the target code/agents branch being reviewed, not failures in the local-review transport itself.