feat(core): make AO-local reviewer backend configurable

[drdreo]

Summary

The AO-local reviewer was hardcoded to shell out to codex, so hosts without Codex/OpenAI credentials (e.g. Claude Code-only setups) hit a 401 Unauthorized on every ao review run and had to paste a --command wrapper into every call. This makes the reviewer backend configurable, defaulting to the agent the user already runs.

Resolution precedence

The reviewer backend resolves in the documented order:

1. --command <shell> flag (unchanged escape hatch)
2. per-project projects.<id>.review (command or agent)
3. global review block (command or agent)
4. the worker agent (defaults.agent) when it has a known reviewer adapter
5. codex (final fallback — preserves existing behavior)

A claude-code worker now reviews with Claude Code by default; codex remains the fallback only when the worker agent has no adapter (e.g. aider/opencode) or already is codex.

Changes

• packages/core/src/types.ts — new ReviewConfig (agent | command, mutually exclusive); review? added to OrchestratorConfig (global) and ProjectConfig (per-project).
• packages/core/src/config.ts — ReviewConfigSchema (strict, mutual-exclusion refinement) wired into both schemas.
• packages/core/src/code-review-manager.ts — runClaudeCodeReview adapter (claude -p <prompt> --permission-mode bypassPermissions --output-format text), a {claude-code, codex} adapter registry, and resolveCodeReviewRunner(...). Resolution happens up front in executeCodeReviewRun so a misconfigured review.agent fails fast (clear error naming valid values) before any run state changes. runReviewer is kept as an explicit override; the --command flag now flows through reviewCommand.
• packages/cli/src/commands/review.ts — run/execute pass reviewCommand instead of constructing the runner; ao review --help documents the config keys.
• Docs/schema — schema/config.schema.json, README.md, agent-orchestrator.yaml.example, and the ao config-help reference all document review.agent / review.command.

Tests

New packages/core/src/__tests__/code-review-backend.test.ts (13 tests) covers the full precedence ladder, review.command → shell runner, fail-fast on unsupported review.agent (global & per-project), no regression when the worker agent is already codex, buildClaudeCodeReviewArgs argv, and parseReviewerOutput round-tripping claude-code-style plain + markdown-fenced JSON.

All existing tests pass (core 1353, cli 905); pnpm build, pnpm typecheck, and pnpm lint are green.