feat: resolve silent button returns with UI alerts, ensure dual action path coexistence, and refactor heuristic review engine

Restrict LLM debate turns to exactly 6 single-round turns, and auto-select/sync first item on category filtering to prevent stuck details panels

Fix HF debate arena fallbacks, score focus 0 in LLM mode, and typewriter concurrent race condition

feat: implement custom product audit in Task B, enforce LLM re-ranking coverage, and fix missing price/category metadata in LLM recommendations

fix: resolve dimension alignment, cache-busting, and streaming early exit bugs

feat: persist active recommendation sub-tab, and dynamically synthesize Judge Agent debate verdict

feat: refine onboarding tour, unify textarea styling, update solution paper tabs, and resolve FastAPI 500 KeyError on custom personas

feat: initial commit