feat(storage): reconstruct portable inputs and exchange legacy bundles - #130

Owning issue

Closes #122

Outcome

Retain and reconstruct explicit market-data input sets independently of mutable storage generations.

• Add storage.SnapshotTable, InputSnapshot, export_snapshot and read_snapshot. Capture frames or protected committed storage keys, caller-supplied effective configuration, source/as-of/unit/revision declarations and existing Specs artifacts.
• Copy immutable relative members, flush and verify staged publication, validate all hashes before parsing, and reject changed destinations. Preserve original keys, nulls, nanoseconds, civil labels and explicit calendar selection.
• Add optional compat.zipline.export_bundle through supported asset/bar/history readers, plus standard-writer make_bundle_ingest for staged migration. Preserve raw daily/minute arrays, supplied factors, dated equities/futures, payout identities/indexes and rename lookup intervals. Reject lossy target price scales before invoking writers.
• Add a runnable native six-table CSV/Parquet reconstruction example and migration/storage/API documentation. Source pruning, replacement and archive relocation do not change reconstructed preparation outputs.

Compatibility and numerical behavior

Existing provider, storage and adjustment defaults remain unchanged. New interfaces are opt-in; Zipline is required only in the optional conversion environment, never for native snapshots. Shared declarations use Specs and the canonical explicit calendar > feed metadata > NYSE precedence; explicit None disables calendars and 24/7 remains scheduled.

Unknown publication times, quote currency, merger consideration and original listing endpoints are not invented. Source history factors retain their own method. Data prepares inputs, not runtime history, analytical cleaning or broker/account execution. A snapshot hash identifies bytes and declarations, not financial equivalence or authentication.

Verification

Final reviewed head: f14de81177a407969a3ec2ede5923035bf5f91bd.

• All 28 hosted checks completed: 27 success plus the configured Python 3.15 exception, whose validator succeeded. All supported Python 3.12-3.14/Linux/macOS/Windows, security, documentation, coverage, ecosystem and installed-bundle gates are green.
• Full local normal lane: 4,531 passed, one optional conversion-environment skip and 292 deselected in 345.63s; coverage 85.75% against the unchanged 85% threshold. Separate ResourceWarning-error lane: 4,531 passed in 362.92s. Ruff, format, ty, strict docs, source-identified rendered docs/internal/external links, builds, actionlint and all hooks passed.
• Seven contract controls pass on the final head with correct and deliberately violated producing-code runs; zero waivers, findings, unreachable or deferred items. The milestone verification check passes.
• Actual CSV frame and committed-key exports reproduce omitted implicit/metadata-only calendars on 50bbcf5 (two failures). The final head passes all 39 focused cases, retaining default NYSE, metadata-only null/24/7 and explicit overrides through shared versioned FeedSpec declarations. Unclocked reference tables remain unclocked.
• Actual standard-writer precision regression fails on d176600 for an incompatible target scale while configured precision passes. The final installed route passes all 13 real standard bundle producer/export/native-factor/re-ingestion cases, including empty selection, stock-dividend recipient identities, daily/minute arrays and dated futures. Interrupted data-flush publication also has failing-base/passing-fix evidence.
• Clean final wheel and source-distribution installs both match all 177 packaged Git source files. Both pass native reconstruction without any Reloaded runtime, actual installed downstream consumers, all 13 compiled Zipline cases and all 39 selected dependency-floor cases. The dedicated hosted installed-conversion job also passes all 13 tests.
• Recovered native outputs pass actual Engineer labeling, Diagnostic forward returns and Backtest contract/position quantities. The Backtest assertion is USD500, not corporate-action account-ledger parity.

Qualified Data 0.2.2.dev30+gf14de8117, Specs 0.2.0, Python 3.12. Native latest uses pandas 3.0.6/Polars 2.0.0; selected floors use pandas 2.3.3/Polars 1.36.1/pandas-market-calendars 5.4.0. Optional Zipline Reloaded 3.2.0 is pinned to cf649ac06fddce76f6caa2403cd8494ac84f72ad; Engineer to c69e7badaac012ad98b1c4b8aa9a67a952b08f5c, Diagnostic 0.1.9, Backtest to 12a2c483f8f1fda54244af187db8bf44e6f85fc9. Upstream deprecations remain visible; ResourceWarnings are errors.

Wheel SHA256: e95ed397a1b3a3580b2b6e013cee20418c045c4750a3b1c2e785226300a6b2bc.
Source-distribution SHA256: ae06fdedf276bac0dd19090d7e4082648794f60f4a15e7eeb48053adac25211e.

AI assistance

Codex assisted with implementation, tests, documentation and evidence review. This submission has been substantively reviewed against its public contracts and independently expected values; it is not an import-only parity claim.

Documentation and release

Existing guides are preserved and extended with runnable migration steps, input requirements and ownership/qualification limits. This is unreleased source work. No release, tag, package publication, documentation deployment or announcement is part of this PR; published Data 0.2.1 remains immutable.