feat: add support for Codex

Description

Codex (https://chatgpt.com/codex). Tier-2 target under epic #66 alongside Gemini CLI.

Codex is a "universal" agent: it reads .agents/skills/ directly, so once the repo is published to the open standard (skills.sh), the workflow skills land in Codex for free. Remaining work is the hook and MCP, plus the subagent-degradation decision.

Scope

• Skills — verified to load from .agents/skills/.
• Hook — port recommend-plugins to Codex hooks (hooks.json or inline [hooks] in config.toml; PreToolUse / UserPromptSubmit events). Caveats: experimental, feature-flagged ([features] codex_hooks = true), not on Windows. The host-aware version of recommend-plugins.sh from PR #220 already detects non-Claude hosts and is a starting point.
• MCP — wire context7 via Codex ~/.codex/config.toml.
• Subagents — Codex has full subagent support (GA March 2026): custom agents as TOML in ~/.codex/agents/, spawn_agent / wait_agent primitives, /agent. Convert our agents/*.md definitions to Codex TOML so the review/plan/research panels run natively.
• CI — loader job asserting skills load in Codex.

Docs: https://developers.openai.com/codex/hooks

Requirements

• Skills load in Codex via .agents/skills/
• recommend-plugins hook ported to Codex hook format (caveats documented)
• context7 MCP wired via config.toml
• Agent definitions converted to Codex TOML (~/.codex/agents/); review/plan panels run natively
• Codex loader CI job added

Additional Context

Part of epic #66. Tier 2 / wave 1 alongside Gemini CLI (#227) and OpenCode (#229).