feat(featuresets): implement time-safe feature engineering layer

INITIAL-4.md

@@ -18,7 +18,13 @@
 ## DOCUMENTATION:
 - Time-series feature engineering best practices
 - scikit-learn transformers/pipelines (if used)
+- [scikit-learn Pipeline Composition](https://scikit-learn.org/stable/modules/compose.html)
+- [MLForecast Feature Engineering](https://www.nixtla.io/blog/automated-time-series-feature-engineering-with-mlforecast?utm_source=chatgpt.com#introduction-to-mlforecast)
+- [sktime Transformations API](https://www.sktime.net/en/stable/api_reference/transformations.html)
 
 ## OTHER CONSIDERATIONS:
 - Feature configs must be persisted per run in the registry.
 - Reproducibility: same config + same data window must be re-runnable.
+- **Imputation Logic**: Define behavior for missing price data (forward-fill) vs missing sales data (zero-fill).
+- **Agent Tooling**: Expose the Feature Pipeline as a tool for PydanticAI to "inspect" the shape of the data before suggesting ModelConfigs.
+- **Computation Overhead**: Evaluate if features should be computed on-the-fly in FastAPI or pre-computed in a materialized view for performance.

INITIAL-5.md

@@ -11,6 +11,13 @@
 - Extensible “Global ML” hook:
   - regression pipeline (scikit-learn)
   - enabled/disabled via feature flags
+- Unified Estimator Pipeline:
+  - Scikit-learn Pipeline incorporating Scaling -> Encoding -> Regressor.
+  - Integration with FeatureEngineeringService for automated lag-injection.
+- Persistence Layer:
+  - Joblib-based serialization including a 'ModelBundle' (Model + Metadata + FeatureHash).
+- Multi-Horizon Support:
+  - Logic for Recursive Forecasting (predicting day-by-day and updating lags).
 
 ## EXAMPLES:
 - `examples/models/baseline_naive.py`
@@ -21,8 +28,11 @@
 ## DOCUMENTATION:
 - scikit-learn estimators + pipelines
 - joblib serialization patterns
+- [scikit-learn Pipeline Composition](https://scikit-learn.org/stable/modules/compose.html)
+- [scikit-learn Glossary](https://scikit-learn.org/stable/glossary.html)
+- [scikit-learn Model Persistence](https://scikit-learn.org/stable/model_persistence.html)
 
 ## OTHER CONSIDERATIONS:
 - No hardcoded horizons: driven by request/config.
 - Determinism: random seed from Settings.
-- Enforce input grain validation (store×product×date).
+- Enforce input grain validation (store×product×date).