The job's applied calibration rounds

Every applied "Apply what I learned" card on the job, newest first (D33): what it replaced (before_ref → after_ref: a filters digest for a search card, a rubric version id for a rubric card), the thumbs it was learned from, the rubric preview's inversions or the filter proposal's pool before/after, and the proposal test that ran on it, if one did. Written by the apply path, never by the model. No pagination: a job has a handful of rounds.