Add a new field to the alert schema to fix a mismatch between the prior schema and the upstream NWS API

docs/integrations/nws.md

@@ -44,6 +44,9 @@ normalizer uses fields under `properties` such as `stationId`, `stationName`,
 `nws_alerts` expects an alerts FeatureCollection. The normalizer uses the
 collection `updated` timestamp, `title`, each feature ID, alert classification
 fields, narrative fields, timing fields, sender fields, and references.
+`properties.onset` and `properties.ends` map to the canonical alert period
+start and end. `properties.expires` maps only to canonical `expires` provider
+metadata and is not treated as the alert period end.
 
 `nws_forecast_hourly` and `nws_forecast_narrative` expect gridpoint forecast
 GeoJSON with `properties.generatedAt`, `properties.updateTime`, elevation,
@@ -98,8 +101,9 @@ unset. Forecast temperatures are converted to Celsius when NWS supplies
 Fahrenheit, and wind speed strings are converted to kilometers per hour.
 
 Alert timing fields are parsed best-effort. Invalid per-alert timestamps are
-left unset rather than failing the whole alert run. Missing alert IDs are
-synthesized from the run snapshot time and array position.
+left unset rather than failing the whole alert run. NWS `ends` is preserved
+separately from `expires`; `expires` does not fall back to `ends`. Missing alert
+IDs are synthesized from the run snapshot time and array position.
 
 Forecast discussion parsing requires an issue time. Weather story entries require
 start time, end time, and update time.