feat(selfhost): complete Codex reviewer modes and fallbacks

[JSONbored]

Summary

Finish Codex reviewer support in the self-host stack so maintainers can choose Claude Code, Codex, Ollama, dual independent reviews, or dual synthesized disposition without breaking existing Claude support.

Context

Claude Code quota can temporarily block reviews, so Codex needs to be a first-class reviewer path. This should preserve repo-scoped Claude instructions and skills while supporting Codex project instructions through AGENTS.md and the existing skills model.

Requirements

• Support Codex as a single reviewer provider.
• Support dual independent reviewer mode, such as Claude Code plus Codex, with separate model and effort settings where supported.
• Support dual synthesized mode where two provider reviews feed one final disposition/comment.
• Preserve Claude Code behavior and repo-scoped CLAUDE.md support.
• Prefer Codex AGENTS.md while falling back to CLAUDE.md when needed.
• Validate provider credentials, CLI availability, model/effort config, and unsupported combinations before running a review.
• Document safe self-host credential mounting for Codex auth.

Deliverables

• Provider selection and validation coverage for Claude Code, Codex, and Ollama.
• Codex runtime tests using realistic CLI JSONL output.
• Dual independent and dual synthesized mode tests.
• Repo-scoped instruction loading tests for AGENTS.md, CLAUDE.md, and skills.
• Troubleshooting docs for Codex auth, quota, and CLI spawn failures.

Acceptance criteria

• A self-hoster can switch from Claude Code to Codex without deleting Claude config.
• Dual provider modes produce the expected comments or synthesized disposition.
• Missing/invalid Codex auth fails clearly before posting partial review output.

[JSONbored]

Closing as completed (2026-07-01 issue audit): Codex reviewer provider + single/consensus/synthesis modes shipped in #1678 (ai-config.ts/ai.ts/ai-review.ts).