fix: correct step format, per-task output, tuple grader returns

FINAL: emit per-task START/END blocks for validator task counting

fix: clamp inference.py scores to (0.001, 0.999) strictly

FIX: Remove incorrect action wrapper in HTTPEnvClient.step() - POST body must be direct ConfigDebugAction, not nested

Fix OpenEnv API step payload format - wrap action in action key

Phase 2: Benchmark Quality Upgrades - Progressive graders, realistic multi-bug tasks, ground_truth exposure