feat(skills): add double-blind-review skill + skill symlinking

[ulises-c]

Summary

Adds agentic-ai/skills/ as a home for cross-agent skills, wires it into the Claude install, and ships the first skill: double-blind-review — two reviewers with different training runs (a Claude subagent and OpenAI Codex) review the same change blind to each other, then cross-examine each other's findings before anything is acted on.

Two logical commits:

1. feat(agentic-ai): symlink cross-agent skills into ~/.claude/skills — install.sh links each skill directory individually (never the parent, so plugin-installed and hand-authored skills survive), -n keeps re-runs from nesting links, and a pre-existing non-symlink is warned about and skipped.
2. feat(skills): add double-blind-review orchestration skill — SKILL.md (the four-round protocol: blind launch → hold → cross-examination → orchestrator verification), round-1/round-2 prompt templates, and a vendored scripts/codex-run.sh so the skill stands alone with no plugin dependency.

Design notes

• Distilled from a live run, where cross-examination surfaced a bypass neither reviewer found alone; the round-2 questions each carry the observed outcome that justifies them.
• No model name pinned — --effort xhigh is the only knob the skill sets; model selection stays with codex's configured default (codex debug models for discovery).
• Standalone by design — the wrapper is vendored from the codex-review plugin (0.4.0), which itself duplicates the file per-plugin for the same reason; the vendored copy's header says to re-vendor rather than hand-edit on divergence.
• Read-only review path — the wrapper defaults to read-only sandbox, prompt text travels as a single argv element from a file (injection-inert), and the skill forbids auto-applying fixes.

Verification

• shellcheck --severity=warning passes on codex-run.sh (also enforced by pre-commit).
• Security review of the branch: zero findings — prompt content cannot be re-expanded by the shell, and the symlink loop cannot be steered outside ~/.claude/skills by crafted names or pre-planted links.
• Privacy scan: no identifying info in the new files (the earlier laza-guidance plugin-path reference was removed before commit).

🤖 Generated with Claude Code