BETWEEN BUILDS

M2 closed: every trial now runs inside a per-task container with memory, CPU and process ceilings and no network interface at all, and a diff is scored by running it - the ground-truth adapter measured at 1.0 and the null adapter at 0.0 rather than asserted. The pinned-image re-mine came back 743 examined and 0 kept again, so the reach limit stands and the negative result gets published. M3 is underway: `assay run`, n trials per task per tool, with pass^n leading.