Add Agent and MCP repository security scanning

[ORANGINGS]

What changed

• Extend skillbom scan to accept local project directories and public https://github.com/owner/repository URLs without requiring an API key.
• Clone public repositories shallowly and non-interactively, disable repository hooks, skip submodules and symbolic links, enforce file/byte limits, and delete temporary checkouts.
• Automatically classify Agent Skill, MCP Server, Agent integration, and generic AI tool repositories.
• Inventory exposed Python and JavaScript/TypeScript MCP tools.
• Detect Node install hooks, remote download-and-execute behavior, Node child-process execution, dynamic JavaScript evaluation, and sensitive credential/browser paths.
• Correlate high-impact code with exposed MCP tools.
• Report capabilities and external domains observed in code but not disclosed in README, SKILL.md, SECURITY.md, or project documentation.
• Add skillbom repo-diff to compare two Git refs for capability drift, newly exposed MCP tools, services, and new high/critical findings.
• Preserve the existing Skill scanner and Policy-as-Code gate through automatic or explicit scan modes.
• Add JSON/SARIF repository reports, release documentation, and focused regression tests.
• Reposition the project explicitly as an Agent Skill/MCP supply-chain scanner rather than a general bug scanner.
• Bump the package to 0.4.0.

Why

SkillBOM's differentiator is evidence-backed detection of undeclared or high-risk behavior in Agent Skills, MCP servers, and AI tool integrations, especially behavior that changes between releases.

The core implementation remains deterministic and does not require a paid AI API. Public repositories are acquired with git clone; semantic LLM analysis and OSV dependency intelligence remain optional future layers.

Validation

• Complete pytest suite passes on Python 3.11, 3.12, and 3.13.
• Ruff passes on the uploaded revision.
• Existing Skill scan and Policy-as-Code positive/negative CI cases pass.
• SARIF generation and GitHub Code Scanning upload pass.
• Focused tests cover public GitHub URL validation, secure clone flags, MCP classification, exposed tool inventory, install-hook execution, child processes, sensitive paths, disclosure gaps, symlink exclusion, repository report drift, and CLI repository scans.