[stefan-jansen]

Summary

Port finite Pipeline feature calculations and dated selection workflows to native Polars without a legacy graph runtime. Add separately named finite methods instead of changing Wilder RSI, recursive EMA, portfolio drawdown or other existing definitions.

• Add sixteen registry-backed finite methods, paired target-price preparation and supplied known event weekday counts. Keep exact typed differences, original windows, target keys, missingness and material method choices.
• Add eligible decision-time cross-sections, Boolean windows and causal feature-panel preparation using supplied Data/caller history and universe facts. Observation cutoff, decision instant and decision-effective universe key are distinct. Missing endpoints never carry stale values forward.
• Add explicit factor-name/factory recipes, native custom factor/filter/classifier examples and an operation reference. No mandatory Zipline, Data, SciPy or store dependency. Regression p-values require the existing optional stats boundary.
• Persist canonical FeedSpec, physical keys, availability, eligibility, lookbacks, price basis and explicit missing-input policy in owned feature artifacts. Preserve omitted versus null calendar selectors and compound entity columns.
• Keep strict complete-input eligibility by default; opt into known-missing kernel rules explicitly. Unknown/future publication and infinite inputs remain ineligible. Reuse existing schedule validation before calculations.

Closes #108.
Closes #109.

Correctness and migration evidence

Actual installed Zipline Reloaded 3.2.0 comparisons include 101 finite/return-factor/default/factory profiles, 19 selection profiles, four original/custom momentum PipelineEngine journeys through the production shifted pricing loader, 55 math/summary/return/event/operator profiles and 28 classifier/filter profiles. Each original/custom journey compares 135 exact output keys, RSI and long/short selections, including known missing quotes and listing/delisting lifetimes. Actual installed ML4T Data history output preserves a causal mean under a future correction and stable ID through split/cash/rename history.

The regression tests also cover large integer/Decimal prices, variance scale, exact nanosecond knowledge boundaries, real missing scheduled bars versus explicit row mode, keyed target alignment, input masks versus final screens, undefined multi-output values, typed Parquet and JSON/YAML handoffs, and explicit pandas presentation. Deliberate producing-code controls pass all seventeen factor/panel acceptance items with no waivers. Independent mathematical expectations supplement legacy comparisons.

One legacy correctness difference is deliberate: MaximumFilter's float-bit integer ordering selects the wrong result for all-negative groups. Native top(1) retains correct numeric ordering and stable ties. Weekday event counts remain distinct from exchange schedules. RateOfChangePercentage is signed percentage points; PercentChange uses a fractional absolute denominator. The migration reference documents these differences.

Verification status

At final source bc2042430ab134bde7fe725878028ad7de3bf224, the full native suite passes 5,555 tests, with 57 configured deselections. Ruff lint/format, ty, executable documentation, strict MkDocs and commit hooks pass. The preceding 4e20f54 has an identical runtime source subtree and passes no-JIT coverage with 95.91% release-critical line coverage and 90.27% branch coverage. The final source additionally repairs actual Windows cp1252 example output and optional-SciPy executable documentation. Isolated installed Python 3.15.0rc1 with SciPy absent passes 458 focused core/compatibility/documentation cases, with 14 explicit optional-dependency skips, not a module-wide skip. Both current security audits pass.

The downloaded hosted candidate 9264c27e6d09c2db2f98dd119bd9571c60e4efd5 has exactly the reviewed source tree. Its wheel SHA256 is a96f8e4a4a015f0960094091d931817562e49344d24390dcdf2e41ad3eacdf4a. All four real PipelineEngine journeys and 55 expression/operator profiles also pass from that installed hosted artifact. Hosted supported-platform checks remain authoritative; merge waits for every applicable final-revision check, including the final Build Package job. Earlier failed and passing receipts retain their exact tested revisions.

This PR does not publish a release, run arbitrary legacy Terms, select source revisions or adjustments, implement Alphalens cleaning/analytics, or claim order/account/performance equivalence. Optional market/loading preparation and final cross-library migration acceptance remain tracked separately by #111 and #112.

AI assistance

Codex assisted with implementation, tests, migration documentation and this PR text. Producing source, regression controls, actual installed workflows and current-head checks are reviewed and verified in this owner session; automated text or import-only checks are not treated as review completion.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.