Update integration handoff and CONUS retraining notes

training/NATIONWIDE_RETRAINING.md

@@ -55,6 +55,23 @@ python training/train_cold_tiled_mainline.py \
   --run-name firewx_fm_conus_seed7
 ```
 
+## Serving-oriented settings in the CONUS template
+
+The CONUS training template differs from the released California checkpoint
+configuration in two serving-oriented ways:
+
+- `positive_tile_placement` is set to `random_containing`, so a fire-positive
+  cell can appear at different positions within a 32-by-32 training tile rather
+  than always being centered.
+- `input_normalization.enabled` is true for continuous channels
+  `0-9, 13, 14, 15`. Weather, canopy, housing, and population channels are
+  normalized from train-split full-map statistics. The dynamic mask, static
+  validity mask, and categorical fuel code are left unnormalized.
+
+These settings are intended for retraining a cleaner CONUS-serving model. They
+do not change the already released California checkpoints, which use native
+source units as documented in `models/wildfire_fm/input_channels.json`.
+
 ## Practical note
 
 The CONUS grid is much larger than the California regional grid. Run cache