EXP-018: feature expansion + per-feature ablation (superset of #2/#3/#4/#22)

[ducroq]

Context

Production LightGBM-Quantile uses 24 features from 4 columns (price_eur_mwh, wind_speed_80m, solar_ghi, load_forecast). energyDataHub already collects ~7 more series that never reach consolidate.py: ned_production, grid_imbalance, cross_border_flows, gas_storage, gas_flows, market_proxies, generation_forecast.

Scope

Consolidate the individual feature issues (#2, #3, #4, #22) into one evaluated experiment, per the provisional position in docs/hypothesis-log.md (2026-08-20).

Two stages:

1. Stage 0 — per-feature-group ablation (cheap, diagnostic): drop lags / rolling / calendar / wind / solar / load, measure MAE + quantile score + coverage.
2. Stage 1 — walk-forward backtest: incumbent (24 features, 3 quantiles) vs candidate (24 + residual load, generation mix, volatility, calendar-gaps), temporal split, 56-day window, vintage-corrected.

Pre-committed gates (ADR-007): calibration — lower AND upper coverage ∈ [0.85, 0.95] AND Winkler ≤ 1.05× incumbent; skill guardrail — paired DM on |y−p50| (α=0.10, HAC=71) not worse.

Prerequisites

• consolidate.py parsers for the 7 unused series (currently hardcoded to 4).
• features_pandas.py + update_shadow.py + online_features.py in lockstep.
• Watch the known freshness skew (consolidate overwrites rows with later vintages → backtest is optimistic).

Compute

LightGBM feature work is CPU-cheap; not gated on GPUs. Three GPU hosts are available over Tailscale (jwasys-b650-eagle-ax.taileb31bb.ts.net, sadaltager, gpu-server.tailce506.ts.net) — held for a possible neural-quantile track, not this.

Status: deferred — pick up next session. See docs/hypothesis-log.md (2026-08-20) and augur#19.