feat(scorer): user-friendly feedback for weak answers and parse failures

fix(demo): load calibration refs and surface parse failures

fix(scorer): brace-balanced JSON parser and tighter prompt for small LLMs

feat: professional UI polish, interviewer-style scoring, null initial radar

fix(space): add accelerate to enable transformers device_map

feat(demo): redesign skill radar for at-a-glance readability

feat: AdaptiveInterviewEnv v1