feat: add get_relevant_context tool — token-budget-aware context selection

[bntvllnt]

Summary

Add a get_relevant_context MCP tool that takes a task description + token budget and returns the optimal subgraph of files to read. Agents waste 60-70% of context on irrelevant files — our graph knows which files matter.

Motivation

• vexp proved 65-74% token reduction using AST-level subgraphs instead of grep
• codebase-memory-mcp reports 99.2% fewer tokens (3,400 vs 412,000) for structural queries
• Context engineering = "finding the smallest possible set of high-signal tokens" (Martin Fowler)
• Our graph already has PageRank — we can rank files by importance and serve the optimal subset within a token budget

Proposed API

tool: get_relevant_context
input: {
  task: string,           // "refactor auth module" or "fix login bug"
  tokenBudget?: number,   // max tokens to return (default: 8000)
  scope?: string          // module or directory to focus on
}
output: {
  files: Array<{ path, relevanceScore, pageRank, summary }>,
  totalTokens: number,
  coverage: string        // "covers 85% of related dependency graph"
}

Approach

1. Use BM25 search to find task-relevant files
2. Expand via dependency graph (dependents + dependencies)
3. Rank by PageRank + relevance score
4. Truncate to fit token budget (highest-value files first)
5. Return file summaries (exports, imports, metrics) not full contents

Priority

Immediate — Token efficiency is the #1 competitive differentiator for AI agent tooling.