v0.3.0 — Codex support

Codex support, from the same set of files.

• SubagentStart hook for Codex, inside the same hooks/briefing-preload
  script: it branches on hook_event_name, so the Claude Code path is
  untouched.
• Codex agents declare skills in a [briefing] table in
  .codex/agents/<name>.toml. Codex's own [[skills.config]] is deliberately
  not read — it means "visible to the agent", not "preloaded".
• Skills resolve against the roots Codex itself searches (.agents/skills in
  project, parent and repo root, then ~/.agents/skills, ~/.codex/skills,
  /etc/codex/skills). plugin:skill works there too — Codex uses the same
  syntax, undocumented but observable.
• Injection goes through additionalContext, with additionalContextLimit: 0
  in hooks.json. At the default threshold Codex replaces large skill bodies
  with a preview and briefs nobody, silently.
• A SubagentStart hook cannot deny a spawn, so a missing skill yields an abort
  instruction to the agent plus a systemMessage to the user — and none of the
  skills that did resolve. No partial briefings.
• .codex-plugin/plugin.json points at the same skills/ and hooks/
  directories. Nothing is duplicated; one hooks.json serves both harnesses,
  verified against Claude Code rather than assumed.
• TOML parsing uses tomllib where available and a regex fallback on Python
  3.10, which CI still covers. A test forces the fallback path explicitly.
• 11 new unit tests for the Codex path, plus end-to-end verification against
  Codex 0.147.0: a declaring agent answered from injected skill content, the
  identical non-declaring agent did not, and a marker confirmed SubagentStart
  reached the hook in both runs.