refactor: replace per-EP feedback loop with measure-improve-verify phases

Instead of improving the tool after every iteration of every EP (which
causes overfitting and bloat), use a three-phase approach:

Phase 1 (measure): Run all EPs once with the original tool
Phase 2 (improve): Analyze ALL results together, find cross-EP patterns,
  make ONE set of concise improvements
Phase 3 (verify): Run all EPs again with the improved tool

This produces targeted, general improvements that work across all EPs
instead of EP-specific bloat.

CLI: benchmark.py measure|improve|verify|full
Co-authored-by: Cursor <cursoragent@cursor.com>

Author: rausingh-rh