v0.2.2: add honest paired benchmark framework

[ctdaniel]

Summary

Adds the v0.2.2 benchmark framework so CQO can be evaluated with reproducible, acceptance-gated evidence instead of guessed token-savings claims.

Benchmark design

• Paired Baseline vs CQO runs
• Shared case_id + pair_id
• Same repository commit / prompt / acceptance criteria
• Two modes:
  • controlled: same starting model/reasoning
  • full-policy: CQO may use its own routing policy
• Efficiency deltas are only calculated when both runs pass acceptance

Observable metrics

• files inspected / changed
• searches
• focused / broad checks
• subagents
• model escalations
• repeated reads
• wall time

Unknown values are omitted rather than guessed.

Added

• benchmarks/README.md
• Chinese benchmark methodology
• result JSON schema
• case-study template
• benchmarks/report.py with text / Markdown / JSON output
• benchmark tests and CI smoke coverage

Principle

Ship the measurement method before publishing savings claims.

Version bumped to 0.2.2.