Block outbound actions with a hook

Keep the rules in a pure function (.claude/hooks/policy.ts) and the Claude Code plumbing in a thin script (outbound-guard.ts), so the policy stays easy to reason about and the script can be run by hand with sample JSON. The hook matches every tool so the audit log is complete, denies with a reason the model can act on, and, the decision that matters most, fails closed: if it can't parse its input it denies, because a hook that crashes with exit code 1 lets the call through.