[dKosarevsky]

Summary

• Reproduce sqlite-utils 4.2.1 flattened-import late-column failure using pinned, SHA-256-verified mledoze/countries data (ODbL), with no invented records.
• Keep dataset downloading explicit; generated JSONL carries an ODbL notice. No data vendoring, runtime dependency changes, or release bump.
• Actual measured reduction: 250 → 176 records, 631,420 → 438,177 bytes; 1683/1683 full-input structural features retained. Check all 855 column definitions, retained values/storage classes, and an independent expected late-column value.
• Compare first-N honestly: 176 initial rows cover 1411/1683 using full-input vocabulary and reproduce the same early failure. Selection is not needed to find this particular failure.
• Record all five timing samples, first observed/subsequent medians, selection overhead, order/batch sensitivity. Generating a fixture for a one-off import is slower. No historical bug, external adoption, or universal speed claims.

Verification

• 125 core unittest tests passed, including 10 offline recipe guards.
• 14 optional real sqlite-utils/dlt integration tests passed.
• Public-data experiment ran five repetitions against fresh SQLite databases; every success/failure assertion passed.
• compileall and git diff --check passed.
• Independent review verified full-vocabulary baseline coverage against pinned per-row ShapeWitness features.

The public-data command is opt-in; existing CI exercises the offline guards. See docs/recipes/public-countries.md and benchmarks/countries-importer.json. External maintainers were not contacted.