Describe diagnostic feedback without misrepresenting channel flow

Clarify that radiation-flux and precipitation channels stay in the fixed recurrent tensor but are zeroed before the next step. Document the zeroing stage in the rollout sequence.

Constraint: Preserve the model's fixed 85-channel recurrent interface

Rejected: Describe diagnostic fields as absent from feedback | the channels remain present even though their values are zeroed

Confidence: high

Scope-risk: narrow

Directive: Distinguish channel presence from value propagation when documenting diagnostic fields

Tested: git diff --cached --check; reviewed the data-details and rollout descriptions together

Not-tested: Inference execution; documentation-only correction

README.md

@@ -145,8 +145,9 @@ FuXi-2.1 operates on 85 channels per time step: 65 pressure-level channels
 (5 variables × 13 levels) and 20 surface channels, plus static forcings supplied as
 constant inputs. Most channels are **prognostic**: they are both input and output
 channels and are fed back during rollout. Radiation fluxes and total precipitation are
-**diagnostic** outputs from a dedicated decoder head; they are predicted but not fed
-back as inputs.
+**diagnostic** outputs from a dedicated decoder head. Their channels remain present in
+the recurrent input to preserve the fixed 85-channel layout, but their values are zeroed
+before feedback so accumulated diagnostic fields are not propagated to the next step.
 
 **Pressure-level variables** (5 vars × 13 levels = 65 channels):
 
@@ -202,8 +203,10 @@ fuxi-2.1/
 
 1. **Load** pre-normalized input (2 frames at t-6h and t)
 2. **Rollout** autoregressively: model outputs next 2-frame state, last frame is the prediction
-3. **Denormalize** each prediction: `output = output * std + mean`, then `expm1` for precipitation
-4. **Save** each step as NetCDF with geographic coordinates
+3. **Prepare recurrence**: retain the prediction for output, then zero radiation-flux and
+   precipitation channels in the recurrent state before the next step
+4. **Denormalize** each prediction: `output = output * std + mean`, then `expm1` for precipitation
+5. **Save** each step as NetCDF with geographic coordinates
 
 The recurrence state stays on GPU throughout the rollout (no CPU round-trip per step).