Total chunk-pointer rows written this run.
Conversations eligible this run (NULL chunked_at or updated_at > chunked_at).
Conversations whose output failed to parse/validate (not marked; retried next run).
Conversations skipped as empty / no-assistant and marked complete.
Conversations whose output parsed to valid JSON with in-range anchors.
Of those, the ones sent to the model (non-empty + at least one assistant turn).