feat: add token savings estimation to tool responses

[halindrome]

Summary

Adds _meta token savings estimation metadata to MCP tool responses, inspired by jcodemunch. This gives users visibility into how many tokens (and dollars) are saved by using structured graph queries instead of reading entire source files.

• New internal/metrics package — TokenMetadata struct, EstimateTokens (char/4 heuristic), CalculateSavings, and Tracker with atomic JSON persistence to ~/.cache/codebase-memory-mcp/savings.json
• Config keys — metrics_enabled (default: true), pricing_model (claude-sonnet/claude-opus/gpt-4o/custom), custom_price_per_token, plus GetFloat64 helper on ConfigStore
• get_code_snippet instrumented — baseline = full source file size via os.Stat (the file the user would have read)
• search_graph instrumented — baseline = sum of unique source file sizes in results (all files the user would scan to find those symbols)
• Fail-open — if metrics are disabled or os.Stat fails, responses fall back to standard jsonResult with no _meta (no regression)
• No new external dependencies — uses character-based heuristic only

Example _meta in response

{
  "_meta": {
    "tokens_saved": 900,
    "baseline_tokens": 1000,
    "response_tokens": 100,
    "cost_avoided": 0.0135,
    "compression_ratio": 0.1
  }
}

Test plan

• go test ./internal/metrics/... -count=1 — 8 tests (calculations, tracker persistence, load existing, missing dir)
• go test ./internal/tools/... -count=1 — includes TestGetCodeSnippet_MetaField and TestSearchGraph_MetaField
• go test ./internal/store/... -count=1 — config constants and GetFloat64
• golangci-lint run ./... — 0 new issues
• No new external dependencies added

🤖 Generated with Claude Code