feat(weather): population-weighted multi-city index + degree-hours / apparent-temperature / dew-point features

[bartzbeielstein]

What

Adds the cheapest, most reliable load-forecasting accuracy levers as opt-in, fail-safe, deterministic weather features. All flags default off, so the pipeline stays byte-identical to the single-point baseline.

New weather/derived.py (pure functions)

• Heating / cooling degree-hours (hdh/cdh) — split the U-shaped temperature–load response into its two arms (ruba23a).
• Apparent temperature (Steadman) and dew point (Magnus) — fold in the humidity driver dry-bulb temperature misses (maia20a).
• population_weighted_average — combine per-city frames into one demand-weighted national index (zimm25a).
• add_derived_weather_features — deterministic orchestrator; raises on a missing source column rather than imputing.

New weather/locations.py

A fixed, deterministic registry of 13 German load centres, weighted by population and deliberately spread across all regions (N/S/E/W) so the national index is not dominated by one climate zone.

Wiring (opt-in, non-breaking)

• get_weather_features gains locations / location_weights (multi-city population weighting) and derived_features / hdh_base / cdh_base / wind_speed_unit.
• ConfigMulti gains use_population_weighted_weather, include_degree_hours, include_apparent_temperature, degree_hours_base_heating/cooling.
• BaseTask.build_exogenous_features wires them through (getattr-guarded for configs predating the fields).

Why

Implements the lecture's "what to adopt next" priorities #1–#2 (degree-hours + apparent temperature + dew point, then population-weighted multi-city temperature) from "Additional exogenous inputs beyond weather and holidays", and the global, population-weighted weather requested on the sf2 roadmap (current weather is a single Open-Meteo point at Dortmund).

Safety (sf2-safe contract)

Keyless (Open-Meteo), deterministic (fixed city list + fixed published weights, no RNG/wall-clock, stable column order), fail-safe (a missing source column / out-of-range humidity / mismatched index raises ValueError, never silent repair). No prohibited dependencies — numpy/pandas only.

Tests & QC

• 46 new tests: pure functions, fail-safe paths, determinism, and the wired get_weather_features path via a monkeypatched Open-Meteo fetch (fully offline).
• Full suite: 2399 passed, coverage 71.9% (gate 68%); weather/derived.py 98%.
• ruff / black / isort clean; REUSE compliant; new modules ty-clean.
• docs/reference/ regenerated (new symbols registered in _quarto.yml); docs-reference-sync gate satisfied.

🤖 Generated with Claude Code