Exercise the replayed read-only SQLite projection and safe-output append without editing ledger files directly. Use the configured Python runtime's sqlite3 standard library with mode=ro to query the projection. It is a snapshot made before this run's safe outputs; accepted writes become durable only after push_ledger_changes succeeds.

1. Query the projection at /tmp/gh-aw/ledgers/smoke/ledger.db for diagnostics. Report malformed or incomplete records, but do not fail solely because of unrelated historical diagnostics.

2. Query state for prior run records. Verify any existing rows have a key matching the JSON workflow_run_id, with record_type equal to repo_memory_ledger_smoke and result equal to passed. An empty table is expected on the first successful run. Also query for the current run:

   SELECT key, value FROM state
   WHERE key = '${{ github.run_id }}'
   LIMIT 1;

3. If no record exists, submit one ledger_append record:

   {
     "ledger": "smoke",
     "operation": "upsert",
     "value": {
       "record_type": "repo_memory_ledger_smoke",
       "workflow_run_id": "${{ github.run_id }}",
       "result": "passed"
     }
   }

4. Do not expect the current run's append to appear in this run's projection. Verify it on the next run; on reruns, reuse an existing current-run row instead of appending a duplicate.

5. Do not inspect or modify ledger shard files directly.

If every check passes, call noop with a brief summary. If any ledger operation fails or a prior persisted record is incorrect, create one issue using the create_issue safe output. Include the failed check, relevant redacted error message, and run URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}. Never include secrets, raw ledger contents, or filesystem paths in the issue.