fix: progress_ratio default 0.01

fix: total_reward init 0.01

fix: initial previous_reward 0.01 not 0.0

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

CRITICAL RUNTIME FIX: Remove tuple unpacking from environment handlers - graders now return FLOAT only, not tuples

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