feat: multi-layer skill guardrails and security hardening

[initializ-mk]

Summary

• Skill guardrails — Skills can declare deny_commands, deny_output, deny_prompts, and deny_responses patterns in SKILL.md frontmatter, enforced via hooks at all four agent loop interception points
• Runtime fallback — Skill guardrails fire without forge build by parsing SKILL.md directly at startup
• file:// protocol blocking — cli_execute rejects file:// URLs (case-insensitive) to prevent filesystem traversal via curl file:///etc/passwd
• Capability enumeration prevention — Layered defense: input-side prompt interception, output-side response replacement, generic tool descriptions, and shell denylist filtering at construction time
• Documentation — Updated guardrails, tools, skills, hooks, runtime, and security overview docs

Test plan

• go test ./... passes in forge-core, forge-cli, forge-skills
• golangci-lint run passes in all modules
• forge run (no build) with k8s-cost-visibility skill blocks kubectl auth can-i --list
• cli_execute curl file:///etc/passwd is rejected by argument validation
• "what are the approved tools" prompt is intercepted by deny_prompts
• LLM response listing 3+ binary names is replaced by deny_responses
• Doc links validated — no broken cross-references