Editorial note: the Research overview still describes the older assessment/witness status. The constructor-only de-scaffolded milestone is closed, and the new learned latch-placement study completed: all four policies improved over initialization, but local arithmetic and workload-aware planning remained better on mean transfer cost; only two policies showed workload-cue dependence. Please treat this as an editorial handoff only: correcting the overview is not a prerequisite for this study, and no shared-site edits or further numerical work are requested from this chat.