History is lake -> Open-Meteo archive; the tail top-up and the recent
observed window are Open-Meteo (archive / forecast past_days); forecasts
stay MET-primary in the fallback composite only. Rationale, measured live
on prod: NASA's MERRA-2 record ran ~1.3F off the ERA5 family with
location-dependent sign and carried no gusts (the Meteostat fill silently
nulled), so serving it beside ERA5 sources skewed percentiles — and the
Open-Meteo archive IS the ERA5 family, so backup- and topup-sourced days
now grade against the climatology without a source seam. NASA fetch code
stays for drift_check only.
