feat: codex adoption audit, sandbox backend status transparency - #224

Summary

Audit of openai/codex (codex-rs) against hawk, per the established adoption-plan pattern:

• New plan doc docs/plans/codex-adoption-plan.md: full capability audit showing every security/runtime idea in codex-rs already has a native hawk implementation — Landlock+seccomp (internal/sandbox/landlock.go, seccomp.go), macOS Seatbelt SBPL generator (seatbelt.go), Windows ACL, bwrap/nsjail/docker selector, egress network proxy (netproxy.go), static code/command verification (code_verifier.go ≈ execpolicy), escalation-token re-validation (≈ shell-escalation)
• Adopted: resolved native sandbox backend now surfaced in the unified status snapshot (permission.sandbox_backend) and text output (Sandbox: <mode> (<backend>)), matching codex's isolation-transparency UX
• Deliberately deferred as RFCs: Code Mode (v8 JS batched tool calls), agent-identity signing, cloud-tasks client

Test plan

• Focused: cmd/internal/status snapshot + sandbox-backend tests
• Full go test ./... green (twice — locally and via pre-push hook)
• make vet, make lint, hawk verify green
• Boundary guards, submodule drift, govulncheck passed in pre-push hooks
• markdownlint-cli2 'docs/plans/codex-adoption-plan.md' clean