feat(review): use GitHub Copilot CLI as reviewer

[ubcent]

Motivation

• Replace the diff-only Claude reviewer with GitHub Copilot CLI so the reviewer can see staged changes in-context (imports/types/related files) while keeping Claude as the arbiter.
• Apply Codex Cloud diffs locally before review so Copilot can operate on the real working tree and then revert changes after review.

Description

• Add src/lib/copilot.ts implementing checkCopilotAvailable() and runCopilotReview() that run copilot in JSON output mode, parse JSONL defensively into ReviewComment[], normalise severity, dedupe comments, and surface CopilotNotFoundError / CopilotReviewError with codes (not_found | review_failed | parse_failed).
• Update the cloud review flow to codex.applyDiff(sessionId) before review, run runCopilotReview() against the service working tree, and always revert local changes with git checkout . after the review; the arbiter still receives the same ReviewComment[] contract.
• Wire Copilot availability check into startup (src/commands/start.ts) so the CLI fails with a clear error if copilot is not present, and update src/lib/review-loop.ts to use runCopilotReview() instead of the Claude reviewer.
• Add CopilotErrorCode to src/types/index.ts and add unit/integration tests (test/unit/copilot.test.ts, updates to test/unit/review-loop.test.ts and tests/integration/start.test.ts) to cover command execution, parsing, and integration points.

Testing

• Ran unit and integration tests with npm test and all tests passed (unit and integration suites included Copilot mocks and exercise the new flow).
• Ran static checks with npm run typecheck and npm run lint and both completed successfully after minor lint fix, and the project built with npm run build without errors.
• Probed the local copilot binary during the run and observed copilot: command not found in this environment, so the tests mock copilot behavior to validate parsing and error handling.

---

Codex Task