Make gpt-5-4-prompting skill model-neutral and multi-agent aware

The bundled gpt-5-4-prompting skill is model-labeled and says it applies to GPT-5.4 workflows, while current Codex installations may use newer models and bounded multi-agent execution.

Suggested update:

• Rename the skill and metadata to a version-neutral Codex prompting name.
• Keep the useful compact contract blocks: task, output contract, grounding, safety, and verification.
• Replace the single-run bias with guidance for bounded parallel lanes when audit or exploration surfaces are independent.
• Require the lead agent to reproduce or verify accepted child-agent findings before reporting them.
• Describe the concrete benefit: less stale model routing, fewer missed independent surfaces, and faster broad reviews without weakening evidence requirements.

The current skill text itself is still useful; this request is about making its scope and orchestration guidance match current Codex capabilities.

Start by locating the bundled gpt-5-4-prompting skill and its metadata, then review every reference to the model-specific name and single-run workflow. Done means the skill is version-neutral, preserves the listed contract blocks, explains bounded independent parallel lanes, and requires lead-agent verification of accepted findings.