Evaluations

Results from make eval — real pipeline runs against the task definitions in evals/tasks/.

Reviewer/QA/Security detection-rate accuracy is not shown here yet — computing it honestly requires a task with a known injected bug to check detection against (Phase 10's failure-injection harness), not invented percentages.