chore: update inference script and core logic for alignment

fix: clamp all scores in reward_engine and environment to be strictly between 0 and 1

fix: full resubmission patch - fix [STEP] format, add close(), expose system_state, fix /state endpoint, improve reward variance

Add openenv-core dependency and server entry point

fix: implement hybrid scoring (semantic + objective) and flexible grader

initial clean push: N-A finalized build