feat: add validate_change MCP tool — pre-flight risk scoring for AI agents

[bntvllnt]

Summary

Add a validate_change MCP tool that lets AI agents submit proposed changes and receive risk scoring before modifying code. This is the missing feedback loop — research proves static analysis feedback reduces AI-generated bugs by 70-80% (arXiv 2508.14419).

Motivation

• AI-generated code has 1.68x more issues per PR than human code (CodeRabbit Dec 2025)
• Amazon lost 6.3M orders from an AI-assisted code change (March 2026)
• No competitor offers pre-flight architectural risk scoring via MCP
• Anthropic's Code Review (March 9, 2026) analyzes "logic relationships between functions, modules, and dependencies" — we should be the infrastructure layer providing this data

Proposed API

tool: validate_change
input: { filePath: string, symbol?: string, changeType: "modify" | "delete" | "add" }
output: {
  riskScore: number (0-1),
  blastRadius: number,
  affectedFiles: string[],
  affectedTests: string[],
  metrics: { pageRank, betweenness, coupling, tension },
  recommendation: "low-risk" | "review-recommended" | "high-risk-review-required",
  hints: string[]
}

Acceptance Criteria

• Agent can query risk before modifying a file
• Returns blast radius, affected tests, and risk level
• Integrates with existing graph metrics (PageRank, betweenness, coupling)
• Works with detect_changes for diff-aware risk scoring
• Tests covering all risk levels

Priority

Immediate — This is the #1 opportunity from competitive analysis. No other MCP tool provides this.