Emit a four-stage deterministic strategy and combination walk-forward evidence

[proerror77]

Parent

#580

What to build

Evaluate selected Factor Bank subsets through the real combination walk-forward path and compile the selected combination into one deterministic strategy artifact with separately typed Signal, Sizing, Risk, and Execution stages.

The first slice uses frozen Factor Bank orientation and normalized equal-absolute weights. No continuous optimization or runtime activation is included.

Acceptance criteria

• The Signal stage binds the Factor Bank revision, selected Factor IDs, orientations, normalized equal-absolute weights, combination rule, and signal threshold policy.
• The Sizing stage defines one deterministic bounded position-mapping rule with frozen parameters.
• The Risk stage defines immutable exposure and loss limits and cannot be changed by MCTS or the Agent during execution.
• The Execution stage binds the exact venue, symbol, order semantics, latency/cost assumptions, and supported event modality.
• All four stages have versioned schemas, content hashes, and explicit parent links and compose into one deterministic strategy identity.
• Every fold fits any required normalization or model state on training rows only and scores the subset on the corresponding validation rows.
• Sealed rows are absent from combination selection, normalization, reward, and evaluation.
• Combination evidence records predictive metrics separately from turnover, post-cost return, drawdown, and per-observation risk metrics.
• The result includes an explicit comparison with both mandatory baseline artifacts under the same frozen protocol.
• The selected subset, weights, strategy identity, and evidence are stable for identical inputs and policies.
• Invalid stage linkage, weight sum, Factor ID, orientation, policy, or metric identity fails closed.
• The artifact has no deployment or order-submission authority at this stage.

Blocked by

#601

[proerror77]

Issue Lifecycle Audit reconciliation (2026-08-06): superseded by merged implementation.

• PR feat(alpha): evaluate deterministic Ridge and CART baselines #663 (merged 2026-08-03): feat(alpha): evaluate deterministic Ridge and CART baselines

Closing as completed — deterministic baseline evaluation for the four-stage pipeline is implemented on main.