Online eval baseline: F1=55.9% P=57.6% R=54.3% (13 PRs)

Offline best: F1=70.9% P=77.8% R=65.1% (v36, 15 mini PRs)
Online baseline: F1=55.9% P=57.6% R=54.3% (v1, 13 random PRs)

Config switched to code-review-benchmark-online, thinking_budget=12000,
max_issues_per_pr=4, all prompt improvements from offline iterations.