v4.13.0

What's new

• Agent benchmarks: New benchmark suite measuring Claude Code token consumption with vs without wonk across 25 code-understanding tasks on 5 real-world repos. Results: 37.4% total token reduction with quality maintained (0.90 vs 0.91 baseline).
• Python class fields: Annotated assignments in classes (e.g. pydantic title: str) are now indexed as variables.
• Granular Python imports: from foo import bar, baz now creates separate references for bar and baz instead of one blob.

Full benchmark report: bench/results/agent_report.md

Full Changelog: v4.12.1...v4.13.0