v3: bug fixes + composed rewards + real training evidence

Finalize SRE scoring engine Models and anti-hacking tests

feat: enhance playground visuals and align with Phase 2 validation spec

fix: clamp all scores strictly to (0.001, 0.999) - fix Phase 2 validation

Fix validator out of range score: replaced all 0.0 rewards with 0.01

Optimize LLama-3-8B hard mode performance by filtering out distractor env logs and making the index extraction explicit

Fix: clamp rewards to (0.01,0.99), run all 3 levels in inference

Update inference.py

Sync inference.py