Refine MRMS validators and reprocess window for pass_2 fix

src/reformatters/noaa/mrms/conus_analysis_hourly/dynamical_dataset.py

@@ -52,12 +52,10 @@ def operational_kubernetes_resources(self, image_tag: str) -> Iterable[CronJob]:
 
     def validators(self) -> Sequence[validation.DataValidator]:
         max_expected_delay = timedelta(hours=3, minutes=30)
-        # Pass 1 and Pass 2 multi-sensor products have additional gauge-collection latency;
-        # radar-only and precipitation_surface (which falls back to radar) are always current.
-        lagged_vars = [
+        # Pass 1 and Pass 2 have gauge-collection latency (~60 min); all other vars are always current.
+        gauge_latency_vars = [
             "precipitation_pass_1_surface",
             "precipitation_pass_2_surface",
-            "categorical_precipitation_type_surface",
         ]
         return (
             partial(
@@ -67,13 +65,21 @@ def validators(self) -> Sequence[validation.DataValidator]:
             partial(
                 validation.check_analysis_recent_nans,
                 max_expected_delay=max_expected_delay,
-                max_nan_percentage=15,
-                exclude_vars=lagged_vars,
+                # precipitation_surface worst-case quarter-sampled NaN is ~30% (most recent
+                # timestamp falls back to radar-only with ~34% structural coverage gaps).
+                # radar_only worst-case quarter-sampled NaN is ~34% (structural coverage gaps).
+                # categorical (PrecipFlag) is radar-derived with no gauge latency, similar coverage to radar_only.
+                max_nan_percentage=35,
+                spatial_sampling="quarter",
+                exclude_vars=gauge_latency_vars,
             ),
             partial(
                 validation.check_analysis_recent_nans,
                 max_expected_delay=max_expected_delay,
-                max_nan_percentage=50,
-                include_vars=lagged_vars,
+                # pass_1 and pass_2 worst-case quarter-sampled NaN is ~38% (1 of 3 checked
+                # timestamps is 100% NaN due to gauge-collection latency, rest are ~6%).
+                max_nan_percentage=40,
+                spatial_sampling="quarter",
+                include_vars=gauge_latency_vars,
             ),
         )

src/reformatters/noaa/mrms/conus_analysis_hourly/region_job.py

@@ -228,7 +228,10 @@ def operational_update_jobs(
     ]:
         existing_ds = xr.open_zarr(primary_store, chunks=None, decode_timedelta=True)
         ds_max_time = existing_ds[append_dim].max().item()
-        append_dim_start = pd.Timestamp(ds_max_time)
+        # Start 3 hours before the dataset's latest timestamp so precipitation_surface
+        # (which falls back to radar-only when pass_2 is unavailable) gets reprocessed
+        # and overwritten with pass_2 data once it becomes available (~60-min latency).
+        append_dim_start = pd.Timestamp(ds_max_time) - pd.Timedelta(hours=3)
 
         append_dim_end = pd.Timestamp.now()
         template_ds = get_template_fn(append_dim_end)