One-off script. Walks Journal/Dreams/dream-manifest-*.json and increments
consolidation_count + sets last_consolidated_at for every (manifest, source)
pair. Idempotent — resets the cursor for any touched sources before
backfilling, so reruns don't double-count.
First run: 7547 embeddings rows updated across 105 unique sources, 416
(source, manifest_date) pairs across all manifests. Distribution now: 422
chunks at count=18 (the dominant dossier-narrative cluster that fills every
NREM in the last 18 days), long tail down to count=1, 12,011 still at 0.
This makes dream_observation.underprocessed_count meaningful — before, all
counts were 0 so the bottom-quartile percentile was 0 and the signal was
degenerate. After, the signal correctly identifies the 12k chunks that have
never been replayed.
