Data export and restore

In an app with no server, the export file is the only copy of a user's data that can outlive the phone, and restore is the only recovery path they can invoke themselves. Both are dangerous in opposite directions: a lossy export silently becomes someone's backup, and a half-applied restore destroys the very database it was meant to protect. This skill governs the portable artifacts; the on-device VACUUM INTO snapshot and WAL rules belong to persistence-drift.

Depth lives in references/formats-and-encoding.md (CSV/JSON/PDF mechanics) and references/restore-and-merge-policy.md (validation ladder, merge strategies).

Non-negotiable rules

1. Backup and export are two different features — never conflate them. Backup is exact, machine-read, round-trippable, and the only thing restore accepts. Export (CSV for a spreadsheet, PDF for a human) is an interop artifact, lossy by design, and must be refused as a restore source. WHY: the moment a CSV is restorable, a format built for readability becomes the file someone's records depend on.
2. Every backup carries an envelope, and restore reads it first. formatVersion, schemaVersion, appVersion, exportedAtUtc, a checksum over the payload, then the payload. Restore verifies the checksum, refuses schemaVersion greater than the app's with a plain message, and migrates an older payload through the same forward-only path as the database (run-migration). WHY: an unversioned file cannot be refused, only misread.
3. Restore is all-or-nothing and never edits the live database in place. Import into a fresh staging database, validate everything, close every handle, then publish by atomic rename with the previous file kept as a rollback until the new one opens cleanly. WHY: a partial restore is worse than a failed one — it leaves a state the user cannot describe and you cannot reproduce.
4. Machine formats carry canonical values only. Integer minor units, SI integers, ISO-8601 UTC instants, stable ids, untranslated enum codes. Never a localized numeral, a formatted date, a currency symbol, or a translated label in a file that will be parsed. Localize only in human-facing exports, which are never re-imported. WHY: ١٢٫٥ and 12.5 and 12,5 are the same quantity and three different parses.
5. Identity is content-assigned, never a rowid. Every exportable row carries a stable id (UUID/ULID) minted at creation. WHY: re-importing the same file must be idempotent — with autoincrement ids it duplicates every record instead.
6. State the merge policy and show it before it runs. Replace-all (wipe then import) and merge-by-id (last-write-wins on a stored updatedAtUtc) are different promises; the confirmation names which one is about to happen and what will be lost. WHY: "Restore" reads as "add my data back" to a user who is about to lose a week.
7. Stream to a temp file, publish by rename. Build the artifact through an IOSink into the app's own temp directory and rename it into place only when the last byte is flushed. WHY: a 100k-row export assembled in a String OOMs a cheap phone, and a half-written file must never be shareable.
8. Nothing leaves the sandbox without an explicit user action, through an injected Gateway. No auto-upload, no background sync, no "helpful" cloud copy. The share sheet and file-save picker sit behind a ShareGateway/FileExportGateway (service-boundary-and-native), faked in tests.
9. CSV is escaped twice: RFC 4180 and against formula injection. Quote any field containing the delimiter, a quote, CR or LF, doubling embedded quotes; and neutralize a leading =, +, -, @, tab or CR so a spreadsheet renders the text instead of executing it. WHY: an unescaped cell is a code-execution vector in the recipient's spreadsheet, and a lost row in yours.
10. Export and restore return typed results, never throw at the UI. Result<ExportArtifact, ExportFailure> / Result<RestoreReport, RestoreFailure> with a distinct failure per refusal reason (checksum, unsupported version, malformed payload, no space, permission denied). WHY: "Restore failed" with no reason is indistinguishable from data loss to the person reading it.
11. A failed export leaves no artifact. Delete the temp file on every failure path and never publish a partial one. WHY: a truncated file that looks like a backup is the failure mode that surfaces months later.
12. Say what an export is. Exports are plaintext and unencrypted unless you built encryption on purpose; the UI must say so at the moment of sharing, and the claim must match the code (see release-and-store-shipping for claim wording).

The envelope

/// The only shape restore accepts. Version fields come FIRST so a truncated or
/// foreign file is rejected by the header rather than by a mid-parse exception.
({
  int formatVersion,        // this envelope's own shape — bump independently of the DB
  int schemaVersion,        // the DB schema the payload was written from
  String appVersion,        // provenance for support, never a compatibility check
  DateTime exportedAtUtc,   // UTC, from the injected Clock
  String payloadSha256,     // checksum over the payload bytes, verified before parsing
  Map<String, Object?> payload,
});

formatVersion and schemaVersion are separate on purpose: the envelope can gain a field without a database migration, and the database can migrate without changing the file shape. A single "version" number conflates them and forces a false choice on the next change. Complete, runnable shape: examples/backup_envelope.dart.

Restore: validate, stage, swap

Future<Result<RestoreReport, RestoreFailure>> restore(File picked) async {
  // 1. Header and integrity BEFORE anything is parsed or opened.
  final envelope = readEnvelope(picked);
  switch (envelope) {
    case Err(:final failure): return Err(failure);           // malformed / bad checksum
    case Ok(:final value) when value.schemaVersion > kSchemaVersion:
      return const Err(RestoreFailure.newerThanApp());       // refuse, never guess
    case Ok():
      break;
  }

  // 2. Import into a STAGING database — the live one is never touched yet.
  final staging = await openStagingDatabase();
  final imported = await staging.transaction(() => _importAll(envelope.value.payload));
  if (imported case Err(:final failure)) {
    await staging.close();
    await deleteStagingFiles();                              // user's data still intact
    return Err(failure);
  }

  // 3. Migrate the staged copy forward through the SAME path as the app database.
  //    4. Close every handle, then publish by rename, keeping the old file until the
  //       new one opens cleanly. See references/restore-and-merge-policy.md.
  return publishStagedDatabase(staging);
}

The staging file, the migration path, and the rename ordering are the whole correctness argument — a restore that writes directly into the open app database has no failure mode that leaves the user where they started.

Choosing a format

Need	Format	Restorable?
Full-fidelity backup the app reads back	Envelope + JSON payload (or a copied DB file)	Yes — the only one
Rows for a spreadsheet or another tool	CSV, RFC 4180, canonical values	No
Something a human reads, prints, or files	PDF, fully localized and formatted	No

Human-facing exports localize at the edge like any other rendering (value-objects-money-and-units, i18n-rtl-l10n); machine formats never do. Copying the live database file directly is a valid backup only through the WAL-safe primitive in persistence-drift — never File.copy on an open database.

Tests that must exist

• Round trip on a hostile fixture. Export → import into an empty database → export again → byte-identical. The fixture carries apostrophes, quotes, commas, embedded newlines, emoji, RTL text with bidi marks, whitespace-only strings, the largest supported number, a DST-ambiguous local instant, and a null-vs-empty pair. 'test1' survives everything and proves nothing.
• Idempotent re-import. Importing the same file twice produces the same state under merge-by-id — not duplicates.
• Every refusal is a test: truncated file, flipped checksum byte, schemaVersion one higher than the app, an older schemaVersion that must migrate, empty payload, and a file that is not this format at all. Each asserts the typed failure and that the live database is byte-unchanged.
• Failure leaves no artifact. Force a mid-write failure; assert no publishable file exists in the export directory.
• CSV escaping. A cell containing =cmd(), one containing a,b"c\nd, and an RTL cell each survive a write→read round trip through a real parser.

Anti-patterns

• CSV as the backup format — no envelope, no types, no versioning; it will be someone's only copy.
• An unversioned export file — the first schema change makes every existing file either unreadable or, worse, silently misread.
• Restoring straight into the live database — no rollback exists; a mid-import failure leaves a state nobody can describe.
• Autoincrement rowids as export identity — re-import duplicates everything.
• Formatting numbers or dates in a machine format — locale-dependent output that round-trips only on the machine that wrote it.
• Building the export in a String/StringBuffer — fine for 100 rows, an OOM crash at 100k, on the cheapest device you support.
• Writing directly to the shareable path — a crash mid-write publishes a truncated file that still looks like a backup.
• Auto-uploading or auto-backing-up "for safety" — it converts an offline app into a data processor, changing the store privacy declaration and breaking the claim.
• Trusting a picked file's extension or name — validate the envelope, not the path.
• A generic "Import failed" message — the user cannot tell a wrong file from lost data; name the reason from the typed failure.
• Silently merging when the user expected replace (or vice versa) — state the policy in the confirmation, in the same words as the code.

Definition of done

• Backup and export are separate paths; export artifacts are refused by restore.
• Every backup carries formatVersion, schemaVersion, appVersion, exportedAtUtc, and a payload checksum, in that header-first order.
• Restore verifies the checksum, refuses a newer schemaVersion, and migrates an older payload through the same forward-only path as the database.
• Restore imports into a staging database and publishes by rename; a failure leaves the live database byte-unchanged.
• Machine formats carry canonical values and stable content ids only.
• The merge-vs-replace policy is explicit in code and named in the confirmation UI.
• Writes stream to a temp file and are published by atomic rename; failures delete the temp file and publish nothing.
• Share/save goes through an injected Gateway on an explicit user action; nothing uploads automatically.
• CSV output is RFC 4180 quoted and formula-injection escaped.
• Export/restore return typed Results with one failure per refusal reason.
• Round-trip, idempotent-re-import, every-refusal, no-partial-artifact, and CSV escaping tests pass on a hostile fixture.

Related skills

• persistence-drift — the WAL-safe on-device snapshot primitive and the DAO layer the importer writes through; never File.copy a live database.
• run-migration — the forward-only path an older payload is migrated through.
• error-handling-typed-results — the Result/Failure spine and the never-lose-data guarantees this feature is the user-facing end of.
• value-objects-money-and-units — canonical storage (minor units, SI, UTC) that makes a machine format round-trip.
• i18n-rtl-l10n — localize-at-render, and why localized numerals never enter a file.
• service-boundary-and-native — the ShareGateway/file-picker seam and its fake.
• release-and-store-shipping — the privacy claims an export feature must not break.
• testing-strategy — hostile fixtures, property/round-trip tests, injected Clock.

References

• RFC 4180 — Common Format and MIME Type for CSV Files: https://www.rfc-editor.org/rfc/rfc4180
• Dart — dart:io IOSink, File.rename: https://api.dart.dev/stable/dart-io/File-class.html
• path_provider — app directories: https://pub.dev/packages/path_provider
• share_plus — share sheet (incl. sharePositionOrigin): https://pub.dev/packages/share_plus
• file_selector — save/open dialogs: https://pub.dev/packages/file_selector
• crypto — SHA-256 for the payload checksum: https://pub.dev/packages/crypto
• OWASP — CSV Injection: https://owasp.org/www-community/attacks/CSV_Injection