feat(adapter/codex): first-class Codex CLI skill runtime

[potiuk]

Goal: Bring OpenAI's Codex CLI to feature parity with Claude Code as a first-class skill runtime for the framework — not just a delegation target. Grounds: RFC-AI-0004 Principle 3 — Vendor neutrality across LLM backends, which names Codex by name as one of the runtimes a framework skill must be expressible against.

Current state: A Claude Code plugin (codex:*) lets Claude delegate to Codex for specific subtasks (rescue, adversarial review). That is not a native Codex skill runtime — a maintainer who has only Codex installed can't use the framework at all today.

What "parity" means:

• Skills under .claude/skills/<name>/SKILL.md are invokable from Codex CLI directly (Codex reads framework skills as native agents)
• The tools/* bridges (gh, jira, gmail, ponymail, vulnogram, cve-org, …) are reachable — these are language-agnostic CLI invocations, so they should port directly
• The setup-isolated-* sandbox model has a Codex-equivalent
• HITL primitives: Codex supports per-action confirmation — verify and document the wiring (RFC-AI-0004 Principle 1)

Suggested approach:

• Map SKILL.md → Codex's AGENTS.md convention (Codex reads AGENTS.md natively)
• Document skill-invocation: how a maintainer runs (e.g.) security-issue-triage end-to-end via Codex
• Map sandbox parity — Codex has its own sandbox model; document the mapping to the framework's bubblewrap + network-allowlist baseline
• The existing codex:* Claude-Code plugin is a reference for how skill ↔ Codex translation can work in the other direction; the native runtime is the inverse

Reference:

• RFC-AI-0004: docs/rfcs/RFC-AI-0004.md
• Codex CLI: https://github.com/openai/codex
• AGENTS.md convention: https://agents.md/
• Existing skill shape: .claude/skills/