[codex] Add calibrated forecast intervals

[amarakramali]

What changed

• calibrate one prediction-interval width per forecast lead using recursive,
  non-overlapping rolling-origin errors
• reserve a chronological calibration window, then refit the exported model on
  all observations after calibration
• add lower and upper interval columns to forecast CSVs and persist calibration
  metadata with the model
• render the interval band in forecast plots and validate paired bounds in the
  Streamlit data loader
• expose coverage and calibration-window controls in both forecast and demo
  commands
• document the method and its time-series coverage limitation

Why

Point forecasts do not communicate how quickly uncertainty changes across a
recursive horizon. A single global residual standard deviation would also hide
the horizon effect and risk training leakage.

This implementation trains the calibration estimator only on data before the
held-out window. It then forecasts each calibration block recursively and uses
the finite-sample split-conformal order statistic independently at every lead.
The production estimator is refitted on all data only after widths have been
measured.

User impact

aep-forecast and aep-demo now produce:

• forecast_xgb_MW
• forecast_xgb_lower_MW
• forecast_xgb_upper_MW

The default target coverage is 90% over a trailing 30-day calibration window.
Coverage remains empirical: temporal dependence and distribution drift prevent
a formal distribution-free guarantee.

Validation

• python -m ruff check src tests streamlit_app.py
• python -m pytest -q — 87 passed
• python -m compileall -q src tests streamlit_app.py
• python -m build
• end-to-end 13-day synthetic demo with 3-hour intervals
• verified lower bound ≤ point forecast ≤ upper bound for every generated row
• inspected saved interval coverage, calibration window, and horizon widths in
  the model artifact