feat: complete reproducible Zipline migration and analytical handoffs - #169

Migration workflows

Closes #160. Closes #156, completing the integration acceptance retained after #164 and the accounting/lifecycle batches #166, #167 and #168.

• Accept Pandas and eager/lazy Polars prices/context at one boundary. Reject ambiguous/null/duplicate event keys and preserve equal-time source ordering.
• Consume verified immutable Data snapshots, including captured Engineer feature declarations. Retain the exact source manifest, consumed table map and normalized identities through native result artifacts; reject changed verified frames before callbacks.
• Add ContinuousFuture/PreparedFutures, continuous_future, current_chain and selected dated-contract reads over Data's causal inputs. The existing native engine still owns orders, fills and accounts. Research history is explicitly registered and cannot become an execution feed.
• Run offline Data preparation -> Engineer features -> Engine -> saved result -> Diagnostic holdings/activity/returns, plus continuous-futures migration and original Zipline 3.2 result conversion. Document immutable source versions, input requirements and actual method differences. Production examples do not require Reloaded packages.
• Exercise the same pinned optional producer/consumer cohort in coverage and clean installed Python 3.12/3.13/3.14 CI on Linux, macOS and Windows. Existing locked/platform/package/security gates remain required. The installed acceptance suite rejects every skipped workflow, and coverage preserves and checks its intentionally installed optional cohort.

Executed evidence

Runtime source: c8cf7b1be2c3d3eb8f532f11a4f2ad58a3b19051. Subsequent commits contain tests, CI, documentation and generated qualification evidence; all 112 runtime modules remain byte-identical.

• Full source suite: 3,270 passed, 9 optional cross-engine skips, 5 benchmark deselections. Global statement/branch coverage 90.67%/81.12%; every critical-module threshold passes.
• Built wheel on Python 3.14 and independently built/installed sdist on Python 3.12: the final selected suite passes 44 installed migration tests per supported environment, no skips, warnings as errors. Dependency consistency checks pass. The five-table snapshot case verifies split quantity, contribution cash, NAV, signals/context and result round-trip.
• Five installed Engine -> strict artifact -> Diagnostic account journeys pass, including first-period fees, all cash, long/short, external contribution and minute round-trip. Both documented offline CLIs execute against the installed package.
• Actual original published Zipline 3.2 captures validate volume and declared first-notice selection, raw fills and additive/multiplicative/no-adjustment histories. Same-session versus preceding-session paired roll marks are explicit choices. No universal arbitrary-chain parity claim is added.
• Failing-base evidence includes six real panel/CSV key workflows before input repair, ignored captured feature metadata before its fix, and three aliased continuous-query failures on the actual pre-fix wheel. Corrected reads reject the wrong query before orders; the tests pass on the fix.
• Fresh current Backtest correctness: 66 synthetic and 17 production pairs pass against hash-verified immutable original outputs. Every production output identity equals the accepted measured output. Performance samples and measurement digest remain byte-for-byte unchanged after an explicit nonmaterial-impact source review; no reference performance engine or licensed authentication was rerun.
• Strict MkDocs, 11,890 rendered links/52 external destinations, installed documentation examples, Ruff, ty, all hooks, reproducible distributions, Twine, locked dependency audit and source security checks pass locally.

The pinned Data and Diagnostic builds declare Python below 3.15; this complete producer/consumer cohort uses Python 3.12 through 3.14. Backtest core Python 3.15 qualification runs separately. Published optional Diagnostic 0.1.4 lacks the new holdings API, so only that newer consumer case requires the holdings capability; the mandatory pinned suite executes it.

The integration sources are recorded in validation/successor-integration-requirements.txt. Diagnostic's source distribution label is 0.1.9; the guide explicitly distinguishes the pinned Git source from the previously published wheel. This PR does not publish or deploy a release and preserves the separately pending 0.1.13 candidate.

Hosted qualification must complete on the final head before ordinary guarded merge.

AI assistance

OpenAI Codex assisted with implementation, documentation, source review, regression reproduction and qualification. The maintainer remains responsible for the changes and reported evidence.