Bulletproof _safe_score for all bad inputs (None, NaN, strings, bool)

Defensive score clamping at all emission points

Widen score epsilon to 0.01 so :.3f formatting stays in (0, 1)

Fix reward.value to always be task_score in (0,1) exclusive

Fix task scores to be strictly in (0, 1) exclusive per OpenEnv spec

Initial submission: SQL Agent OpenEnv for Meta+HF hackathon