feat: refactor Task 5 pipeline and enhance sandbox security

feat: enhance sandbox security and update docs

fix: implement auto-recovery fallback for sandbox MemoryError

fix: enforce csv module and robust keyword extraction

fix: genuine LLM gen for Task 5, no hardcoding, no numpy crashes

fix: deterministic Task 5 agent - fast and reliable scoring

feat: add session audit history table to Gradio UI

feat/fix: update app.py for Task 5 without repetition penalty

fix: disable LoRA for Task 5 code gen, add repetition penalty

feat: add Task 5 NDA Data Review with Python sandbox execution

feat: add Gradio demo UI with live LoRA inference and deterministic grading

Final OpenEnv submission - all requirements met