Test what compaction forgets

The experiment is three pieces that never call each other directly: a transcript generator that plants each gold fact exactly once, a compactor (one claude -p call per summary), and a checker that asks all ten questions against one context and scores the answers in code. The decision that matters most is the two controls. The full transcript must score 10/10 and an empty context 0/10. If either fails, the matcher or the question prompt is wrong, and no summary result means anything.