ci: add codex workflows for backport and fixing CI

[jackye1995]

Introduce 2 CodeX workflows that could be commonly used:

1. patch a merged PR to a specific release branch
2. fix a CI workflow that is currently breaking main branch

[github-actions]

PR Review

P0 Security Concern

--dangerously-bypass-approvals-and-sandbox flag exposes the runner to risk.

Both workflows use codex exec --dangerously-bypass-approvals-and-sandbox, which allows the AI agent to execute arbitrary commands without any sandboxing. Combined with workflow_dispatch inputs that could be manipulated, this creates risk:

1. The PR_URL, RELEASE_BRANCH, WORKFLOW_RUN_URL, and GUIDELINES inputs are directly interpolated into the prompt, which is then executed by an unsandboxed agent. While the prompt instructs the agent to behave in certain ways, the agent may interpret malicious inputs unpredictably.

2. The agent has write access to the repository (via LANCE_RELEASE_TOKEN) with no guardrails.

Recommendation: Consider if sandboxing can be enabled with explicit tool allowlists, or restrict who can trigger these workflows (if: github.actor == 'maintainer-list').

---

P1 Issues

1. Missing branch protection validation - The backport workflow allows targeting any branch via release_branch input. Consider validating that the target branch matches an expected pattern (e.g., release/*) to prevent accidental pushes to protected branches.

2. Potential infinite loop on complex conflicts - The backport prompt says to "try to resolve" conflicts and "if conflicts are too complex to resolve automatically, abort." This is vague guidance for an AI agent. Consider adding a concrete limit (e.g., "If more than 5 files have conflicts, abort").

3. No cleanup on failure - If the workflows fail partway through, branches may be left behind (backport/pr-* or codex/fix-ci-*). Consider adding cleanup in a finally block or post-step.

---

The overall structure and prompt engineering are reasonable for the intended use case.