modeld: cached-output standby while gear is in park

Narrow re-introduction of the power-save mode: when carState reports
gearShifter == park, serve the last rendered model_output instead of
running GPU inference. First cycle (or before carState is flowing)
still runs one real inference so we have something to cache and serve.
Out-of-park returns to per-frame inference immediately.

Avoids the pitfall of the earlier variable-rate path: this doesn't
change the publish rate on modelV2/cameraOdometry (we still publish
every frame), so downstream freq_ok / valid checks in calibrationd /
locationd / paramsd stay happy. Only the GPU inference is skipped.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

selfdrive/modeld/modeld.py

@@ -183,6 +183,10 @@ def main(demo=False):
   model_transform_main = np.zeros((3, 3), dtype=np.float32)
   model_transform_extra = np.zeros((3, 3), dtype=np.float32)
   live_calib_seen = False
+  # CLEARPILOT: cache last model output to serve while gear is in park — saves
+  # GPU inference cost while still giving downstream a constant publish rate so
+  # freq_ok / valid checks don't cascade.
+  last_model_output = None
   nav_features = np.zeros(ModelConstants.NAV_FEATURE_LEN, dtype=np.float32)
   nav_instructions = np.zeros(ModelConstants.NAV_INSTRUCTION_LEN, dtype=np.float32)
   buf_main, buf_extra = None, None
@@ -314,12 +318,21 @@ def main(demo=False):
       **({'radar_tracks': radar_tracks,} if DISABLE_RADAR else {}),
     }
 
-    mt1 = time.perf_counter()
-    model_output = model.run(buf_main, buf_extra, model_transform_main, model_transform_extra, inputs, prepare_only)
-    mt2 = time.perf_counter()
-    model_execution_time = mt2 - mt1
+    # CLEARPILOT: in park, serve the cached last model output instead of running
+    # GPU inference. First cycle (no cache yet) still runs once so we have
+    # something to serve. Out-of-park resumes fresh inference every frame.
+    parked = sm['carState'].gearShifter == car.CarState.GearShifter.park
+    if parked and last_model_output is not None:
+      model_output = last_model_output
+      model_execution_time = 0.0
+    else:
+      mt1 = time.perf_counter()
+      model_output = model.run(buf_main, buf_extra, model_transform_main, model_transform_extra, inputs, prepare_only)
+      mt2 = time.perf_counter()
+      model_execution_time = mt2 - mt1
 
     if model_output is not None:
+      last_model_output = model_output
       modelv2_send = messaging.new_message('modelV2')
       posenet_send = messaging.new_message('cameraOdometry')
       fill_model_msg(modelv2_send, model_output, publish_state, meta_main.frame_id, meta_extra.frame_id, frame_id, frame_drop_ratio,