Run a memory ablation on synthetic matters

The harness follows Spencer's order strictly: script, then text, then questions. Every gold answer is read from the script, so a rendering that drops a fact shows up as a verifier miss, not as a silently "correct" question. The four arms share one answer prompt and one model; only the context changes, so any gap between arms is about memory. The decision that matters most for honest numbers is the unit of resampling: questions within a matter are correlated, so intervals are bootstrapped over matters, which makes them wide with six matters. That is the right answer, not a flaw.