fix: batch of 5 stability fixes

feat(spark): update judge with arc-based criteria and challenge support

feat(spark): add inter-round commentary API endpoint

feat(spark): swap AI arguments to Claude Sonnet 4.6 with 150-word cap

fix: judge now scores on debate technique -- rebuttals, pressure, consistency

fix: judge evaluates delivery/reasoning/engagement, not crowd appeal

fix: simplify judge and stop AI opponent from fabricating anecdotes

fix: judge now penalizes repetitive emotional appeals and rewards specificity

fix: ensure "Different topics" generates 3 fully new topics

refactor: rewrite judge prompt for crowd-persuasion scoring

fix: completely rewrite fact-checker to only flag objectively false claims

fix: fact-checker must not cite single org as sole authority

fix: fact-checker bias and AI word limit

fix: AI opponent takes opposite side based on spectrum + judge anti-style bias

fix: judge must track cross-round references and contradictions

fix: add explicit bias prevention rules to judge prompt

fix: overhaul judge and AI opponent for substantive debate

fix: make debate topics substantive and real-world

feat(spark): DB queries, API routes, and UI components for spectrum redesign

feat(spark): critical reliability - disconnect detection, API error handling, judge timeout

feat: redesign judge to be qualitative instead of numeric categories

fix: use lazy OpenAI client init to avoid build-time crash

feat: add Spark Micro-Debate Arena