[stefan-jansen]

Native outcome preparation

Closes #110. PR #113 is merged as 9dcd4b0c97f0776d8ac8a40dc33c6ac767e93ed8.
The branch now includes its actual main ancestry and is ready for final hosted qualification.
Both workflows reuse one precise arithmetic kernel; this is not a replacement precision PR.

Add ml4t.engineer.labeling.forward_returns, returning a normal keyed long Polars table.
Keep existing fixed-label signatures/defaults and the sole Alphalens facade in Diagnostic.
Engineer does not clean/bucket factors, fill quotes, acquire data or reconstruct adjustments.

The producer supports multiple named horizons, sparse exact entry keys, observation/expected-bar/
session/elapsed units, per-asset/common clocks, cumulative/exit-bar returns, exact nanosecond
targets and tolerance, explicit tail/missing-price policies, known-at availability and price basis.
Original row/quote masks from externally prepared research views survive entry/reference/exit
selection. Inserted fill rows cannot clear named-calendar completeness checks.

Reuse the existing schedule/grid adapters and future-price lookup. Extend owned LabelSpec fields
with shared FeedSpec serialization, preserving calendar omission/null and old wire keys when unused.
No new shared contract, graph runtime, adjustment service or Alphalens namespace.

Reproductions and implementation

• Original main c69e7badaac012ad98b1c4b8aa9a67a952b08f5c lacks native multi-horizon production.
  Existing public workflows reject a declared 1ns grid and lose outcome declarations at persistence.
• Public elapsed-window cases reproduced microsecond rounding (102 instead of the correct101.5).
  Shared future-price lookup reproduced ns/us join mismatch; it now retains exact durations/keys.
• Existing UInt64 fixed labels reproduced returns around1.8e17 for a decline that should be-.1.
  Shared integer-safe arithmetic repairs this without changing established floating definitions.
• An actual installed outcome wheel reproduced a positive horizon wrapping into the past at the
  nanosecond limit. Out-of-range inputs/targets now fail atomically.
• Decimal price production reproduced zero tiny returns. Preserve scale through subtraction,
  then convert to Float64. Decimal return output dtype changes explicitly; original price columns
  and existing Float32/Float64 definitions stay unchanged. Opposite-sign limits are covered.
• Real early-close, holiday, DST and overnight cases distinguish expected bars, session deadlines
  and physical targets. A covered session with no full eligible bar remains a missing endpoint,
  not a skipped session or an outside-tail assertion.

Qualification so far

Current implementation head: 3c916aeeabd3742f4ae833d019a42e069e82c0ae.
It has exactly the previously qualified 8732c89 tree
7705f10d88fc222ea81c8a80a57ab5e3ffdc4f4a after resolving the two prerequisite-history
conflicts. No source behavior changed during that integration.

• Focused refreshed arithmetic/outcome/artifact/time-horizon cases: 456 passed.
• Fresh full configured suite at committed 3c916ae: 5,094 passed, 57 configured performance deselections, 17 warnings in 207.06 seconds.
  The actual critical coverage gate passes with 96.05% line and 90.31% branch coverage.
• Ruff lint/format, ty, actual commit hooks, executable guide workflows and strict MkDocs pass.
• Exact 3c916ae wheel SHA256 930a46661b34ea9b00306eefbf7e8b94a1f72680150f6b30b071b0d70d8a2d88;
  sdist SHA256 018d5eb2d43d7e5c4668f10aa9bd85bd67b301e82b47e9ed8ae88041fc85c2d6.
• Actual isolated producer -> Parquet/LabelSpec JSON -> Diagnostic prepared panel passes ordinary
  cumulative and reference-key exit-bar values against the receiver's public price route. Adjacent
  nanosecond keys/exits, string and compound stable IDs, availability and persisted timedelta replay
  survive the native handoff. All 201 Engineer and 289 Diagnostic installed package files were
  byte verified; five receiving implementations independently match source 643b94b. Both wheels
  are unpublished qualification artifacts, not a claim about published Diagnostic 0.1.9.
• The issue verification gate covers all nine obligations with zero waivers or deferrals.
  Producing-code mutations detect contaminated asset keys, wrong clocks, filled-grid acceptance,
  quote filling, lost availability, ignored calendars, wrong endpoint method, rounded durations,
  numeric scale loss, calendar omission loss and non-replayable horizon serialization.
• The new producer is included in the existing critical coverage gate. No thresholds, failure
  checks, supported-platform matrix or security gate were waived.

Owner review of the actual 15-file final diff found no remaining source finding. Before merge,
every applicable final-head hosted source/artifact/platform/security/docs gate must pass. Fresh
full source and rebuilt installed-artifact convergence pass at 3c916ae; earlier unchanged-tree
evidence remains scoped to its stated build.
No release, tag or deployment is requested.

AI assistance: Codex implemented code, tests and documentation, reproduced the numerical defects,
and ran/reviewed the checks above. No independent human review or publication is claimed.

[Copilot]

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