Merge PR#2 features: skip_post mode, reasoning_effort, online eval, lower thresholds

- Add skip_post field to ReviewRequest for eval-only reviews (no GitHub posting)
- Add reasoning_effort="high" for non-tool OpenAI calls (GPT-5 workaround)
- Lower suppressed category thresholds from 0.99 to 0.70
- Add online_eval.py pipeline (discover, dispatch, judge, aggregate)
- Add eval_results.md documenting 4 online eval runs (best F1=0.418)
- Update .gitignore for output files and eval artifacts