fix: re-judge experiment rollouts after credit-exhaustion; retry logic in batched + single-criterion judge

feat: V₀=0.5 baseline + det ceilings, prompt-pill UI, 60 experiment rollouts

feat: per-turn state value + turn score on Demo (state_v1 trajectories baked into 228 rollouts)

restore spivak PDF + 5 spivak tasks

deploy