Select regressor beliefs per (event_start, regressor) in forecasting covariate assembly

[Copilot]

Forecast covariate preparation could drop valid regressor values when multiple regressors had beliefs for the same event_start at different belief_times. Selection was done per joined row (event_start), not per regressor, causing false missing data downstream.

• Covariate selection logic

  • Added per-regressor selection in BasePipeline.split_data_all_beliefs via a helper that picks one value per (event_start, regressor).
  • Applied this consistently across:
    • past regressors (latest belief per event),
    • realized future regressors (closest realized belief per event),
    • forecast-window future regressors (latest admissible forecast belief per event).
  • This preserves independently known regressor values while still producing one wide covariate row per event_start.

• Safety guard

  • Added explicit validation for selection mode (latest / closest) to prevent silent misuse.

• Regression coverage

  • Added a focused test that reproduces mixed-belief-time regressor rows for the same event and asserts both regressor values are retained in the assembled future covariates.

# Before (row-level collapse):
event_start=10:00, belief_time=09:45, regressor_A=NaN, regressor_B=7.0

# After (per-regressor selection):
event_start=10:00, regressor_A=5.0, regressor_B=7.0