Clarify WildFIRE-FM input channel contract

models/wildfire_fm/README.md

@@ -24,6 +24,9 @@ CONUS retraining recipe is provided in `../../training/NATIONWIDE_RETRAINING.md`
 
 The checkpoints expect a fixed 16-channel tensor in `[channel, y, x]` order.
 A machine-readable copy is stored in `input_channels.json`.
+Channel order matters. The released checkpoints use native source units with no
+mean/std normalization or min/max scaling. Invalid or missing values are
+zero-filled, with validity-mask channels carried in the input tensor.
 
 | Channels | Names | Source |
 |---:|---|---|
@@ -36,6 +39,11 @@ FIRMS detections are used to derive the occupancy target, not as an input
 channel. WFIGS and MTBS are supporting event-level resources and are not part of
 the pretrained occupancy input.
 
+For HRRR, `t2m` and `d2m` are 2 m fields in Kelvin; `u10` and `v10` are 10 m
+wind components in m/s. The `cape` channel is selected from the HRRR surface
+instantaneous group with `typeOfLevel=surface` and `stepType=instant`; it is not
+the most-unstable or layer CAPE field.
+
 ## Custom spatial serving
 
 The checkpoint remains a native-grid model. For web and client migration, use