Add self-improving-operator curated skill

[wd041216-bit]

Summary

This PR adds a new curated skill, self-improving-operator, under skills/.curated/.

The skill is designed for cases where the user wants Codex to keep improving a project instead of stopping after a single narrow edit. It encodes a practical loop: inspect the real current state, choose the highest-leverage unfinished gap, make one bounded improvement, verify with real evidence, update handoff material, and continue until the next meaningful bar is reached.

Why this is useful

• Gives Codex a reusable pattern for proactive ownership without encouraging reckless scope creep
• Emphasizes verification, documentation, and handoff quality
• Composes cleanly with other skills when a domain-specific subtask is the better immediate fit

Files added

• skills/.curated/self-improving-operator/SKILL.md
• skills/.curated/self-improving-operator/agents/openai.yaml
• skills/.curated/self-improving-operator/LICENSE.txt

Validation

• Verified the skill structure locally in the standalone development repository: https://github.com/wd041216-bit/codex-self-improving-operator
• Confirmed the upstream diff adds only the new curated skill directory

[wd041216-bit]

Refreshed this PR to the autonomy kernel v2 package. The curated skill now ships a real .operator/ runtime, plan-to-backlog decomposition, checkpoint/resume flow, generated references, and tighter signal hygiene so stale or non-actionable TODO mentions do not linger in the queue. The canonical repo and generated bundle are back in sync, and the source repo validation is green.