fix(phase2): return score-only grader payloads to avoid out-of-range numeric fields

fix(grader): ensure minimum mathematically bounded score of 0.01 directly so that .2f float representations never evaluate down to literal 0.00 strings inside inference logs

fix(eval): mathematically bound all precision/recall metrics via laplace smoothing to naturally yield strict (0, 1) scores without artificial clipping

fix(env): remove intermediate step rewards to ensure total episode reward strictly scales in (0, 1) bounds

fix(grader): strictly bound task scores between (0, 1) for hackathon validation

fix: clip grader score to strictly (0, 1) for validator compliance

Fix score clipping logic to enforce Phase 2 strict boundary (0,1)

feat: implement FastAPI server, Streamlit dashboard, and unit tests for SilentFailureDetector environment

Include OpenEnv MVP implementation and inference baseline fixes