feat: add v0.5 agent integration workflow

[DOTfei]

Summary

Vibe Code Guard v0.5 makes the repository usable by external coding agents
such as Codex, Claude Code, Gemini CLI, and Cursor agents.

The product boundary is explicit:

Coding agent = project intelligence, explanation, and authorized code changes
Vibe Code Guard = toolchain, workflow, findings, lifecycle, release gate, Dashboard
Upstream scanners = actual detection engines

Agent-first experience

An agent can receive this repository and the instruction:

Install and use Vibe Code Guard from https://github.com/DOTfei/vibe-code-guard,
then security audit this project.

The new flow provides:

• root AGENTS.md and Claude-compatible CLAUDE.md instructions;
• one canonical workflow in docs/agent-integration.md;
• safe installation/bootstrap through ./install.sh;
• a repository-controlled config/toolchain.json manifest;
• structured doctor --json status and explicit READY/DEGRADED/BROKEN states;
• canonical audit profiles: auto, quick, full, and release;
• stack-aware use of the existing deterministic orchestrator;
• structured run output with correlated findings, release gate, and Dashboard data;
• vibe-code-guard dashboard with a 127.0.0.1-only binding;
• safe project config through .vibe-code-guard.json;
• launcher path hints when shell PATH is intentionally left unchanged;
• Vibe Code Guard-only update and uninstall behavior; and
• manual fallback documentation for humans.

Safety and ownership boundaries

• Upstream scanners remain independently installed and are not bundled,
  modified, or relicensed.
• The installer uses fixed official Homebrew/pipx actions, does not use
  curl | sh, does not change shell startup files, and does not disable OS
  security controls.
• Healthy existing scanners are not reinstalled or downgraded.
• Uninstall removes only launchers and metadata recorded as Vibe Code Guard-owned;
  upstream scanners, databases, and templates are preserved.
• Runtime scanning is localhost/local-Docker by default. Exact non-local targets
  require explicit authorization; public targets are never inferred as safe.
• Project config accepts only profile, safe relative ignored paths, and runtime
  target declarations. It cannot execute arbitrary shell commands.
• v0.4 AI Review remains optional and disabled by default. No real external AI
  provider, API key, cloud upload, GitHub PR review, automatic fixing, Strix
  integration, or new scanner is added in v0.5.

Validation

• npm test: 50 passing
• Node syntax checks: passing
• bash -n install.sh: passing
• git diff --check: passing
• Gitleaks: no leaks found
• ./install.sh --dry-run --json: passing; planned actions are structured and
  no machine changes were made
• fresh-machine simulation: missing tools/installers, non-default PATH,
  incompatible version, healthy newer version, root/symlink escape, command
  conflict, and upstream-preserving uninstall all covered
• isolated --yes bootstrap: DEGRADED was reported accurately, absolute
  launcher path returned, launcher executed, and uninstall removed only VCG files
• update check and uninstall dry-run: passing in isolated temporary toolkit home
• quick audit integration: scanner errors, SKIPPED states, correlated counts,
  deterministic release gate, and structured error reasons were verified
• Dashboard start/restart smoke test: READY, bound to 127.0.0.1, safe port
  selection, structured dashboardUrl, isolated temporary data, then stopped
  and cleaned up
• global security-tools doctor: HEALTHY
• global security-tools self-test --json: 7/8 PASS, 1 DEGRADED, 0 FAIL

The self-test degradation is the pre-existing global OSV-Scanner external query
condition (exit 127, no results), not a v0.5 repository failure. It is reported
accurately and was not changed or hidden.

The final review fixes in commit 35abe40 add symlink-parent protection,
version-range validation, IPv6 loopback support, rejection of 0.0.0.0, explicit
manifest doctor/self-test mappings, scanner error/skip fields, prerequisite
documentation, and a safe launcher path hint.

Compatibility

Existing v0.1 orchestration, v0.2 Unified Findings, v0.3 Correlation/Lifecycle,
v0.4 optional AI Review, run history, Dashboard APIs, and third-party notices
remain in place. The v0.5 CLI calls the existing deterministic audit workflow;
it does not replace scanner adapters or change deterministic authority.

Known limitations

• The installer plans supported upstream installs through official channels but
  does not silently approve them; --yes is required to apply the plan.
• A pre-existing unrelated global security-check is preserved rather than
  overwritten. The unambiguous vibe-code-guard command remains available.
• Scanner database/rule/template freshness remains governed by the global
  security-toolkit lifecycle.
• macOS Apple Silicon is the primary supported path; macOS Intel and Linux are
  partial, and Windows/WSL is not yet supported by the installer.
• No autonomous code changes are performed by Vibe Code Guard. The external
  coding agent must obtain user authorization, fix, rescan, and report.

Recommendation

Final productization review: no remaining blocking v0.5 defect is known from
the performed review. Recommend MERGE after human review. Keep this PR Draft
until the owner explicitly marks it ready; do not merge automatically.