- C5A objects table: object_id nullable + reason_code column so HUMAN_REVIEW/
SECRET_DETECTED object changes (object_id=None) are persisted, not silently
dropped (was: object_id TEXT NOT NULL, no reason_code field)
- C5B poll_once: content_before now read from parent_sha (state BEFORE the
change) instead of sha, so MODIFIED changes classify as CONTENT_UPDATE
instead of being misclassified (Checker-Befund)
- test_c5b: add test_modified_content_reads_parent regression test
C5B 40/40, C5A 25/25, real-repo dry run: 233 object changes (109 IN_SCOPE
with valid id, 124 HUMAN_REVIEW), idempotent.
