- **Weakest point in the workflow:** The CSV schema mismatch causes every current fixture import to fail with `KeyError: record_id` before row validation. Separately, committing database changes before writing the local checkpoint creates a failure window: records may be stored while the checkpoint still indicates unfinished work. Enabled automatic retries violate the operator-approval requirement. - **Concrete safeguarding recommendation:** Check CSV headers before processing rows. Explicitly map the renamed `id` column to `record_id`; reject missing or ambiguous headers. Validate each row and assign a report outcome: inserted, already present with the same label, conflicting label requiring review, or invalid. Use `INSERT … ON CONFLICT (record_id) DO NOTHING RETURNING record_id`; when nothing is inserted, read and compare the existing label without updating it. Store row outcomes and run progress in PostgreSQL in the same transaction as inserts. Stop on HTTP 401 and disable automatic replay. - **Simplified logic alternative:** Operator starts run → fetch CSV → check headers → validate rows → insert or classify duplicates → commit records and row outcomes together → generate the count report from committed outcomes. Use a stable run identifier and input row number so repeated IDs still receive separate report entries. A separate local checkpoint is unnecessary. - **Typical error that occurs later:** A timeout after a successful commit is mistaken for a failed insert. Replaying with plain `INSERT` then triggers a primary-key violation, potentially interrupting processing of remaining valid rows. After a commit timeout, mark the outcome as uncertain and reconcile committed database outcomes before offering an operator-approved replay. - **An immediately implementable fix:** Change the importer’s explicit column mapping to `id → record_id`, check headers before entering the row loop, and disable automatic retries in the fixture setup. Run the corrected fixture only after operator approval. These changes address the current blocker; transactionally stored row outcomes address the remaining checkpoint failure window.