feat: 7 tasks, context-gated rewards, live dashboard, heuristic baseline, holistic grading

fix: add score= to [END] clamped to [0.001,0.999], try/finally for [END], safe LLM params

feat: add root Dockerfile, inference.py, and updated env files