locationd: ignore GPS as Kalman input; expand park-cached-output to onroad consumers

locationd: ignore gpsLocation observations entirely (clearpilot_disable_gps
const at the top of handle_gps; falls through to determine_gps_mode's
no-GPS path). gpsd.py keeps publishing real GPS for UI / dashcam / clock
/ night-mode — only locationd ignores it. The previous gpsd.py path was
hard-coding vNED=[0,0,0] while the car was moving 28 m/s, feeding the
Kalman contradictory GPS-vs-IMU velocity observations that propagated into
latcontrol_torque through liveLocationKalman.angularVelocityCalibrated and
caused a real "drift right on straight roads" symptom.

controlsd: short-circuit state_control() while parked. Skips LaC/LoC PID,
MPC, model_v2 reads, lane-change logic — all wasted work when the car
isn't moving. Returns the same enabled=False/latActive=False/longActive=False
CC the original code would have produced, so publish_logs runs unchanged
and card.controls_update keeps the sendcan / tester-present heartbeat
flowing at 100Hz. controlsd CPU dropped from ~59% to ~3% in park.

controlsd: also wire self.FPCC.noLatLaneChange = True/False in the
existing lane-change suppression branch. The Hyundai carcontroller already
reads off frogpilot_variables.no_lat_lane_change for the no-steer signal,
but the UI's distinctive yellow CHANGE_LANE_PATH_COLOR (onroad.cc:901)
reads from frogpilotCarControl.NoLatLaneChange — which nobody was setting,
so the yellow line never appeared during lane-change suppression.

plannerd: skip longitudinal_planner.update + publish + publish_ui_plan
while parked. Downstream controlsd already short-circuits in park, so
longitudinalPlan / uiPlan staleness is fine.

frogpilot_process: same idea — skip frogpilot_planner.update + publish
while parked.

dmonitoringmodeld: mirror the cached-output trick from modeld. Add carState
to the SubMaster, cache the last (model_output, dsp_execution_time) tuple,
and serve it on every frame while gear is in park instead of running DSP
inference on the driver camera.

Together with the existing modeld park-cached-output, the heavy onroad
processing pipeline (modeld → plannerd → controlsd → carcontroller) now
all idles in park and only fires when the car leaves park.

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

selfdrive/modeld/dmonitoringmodeld.py

@@ -7,7 +7,7 @@ import ctypes
 import numpy as np
 from pathlib import Path
 
-from cereal import messaging
+from cereal import car, messaging
 from cereal.messaging import PubMaster, SubMaster
 from cereal.visionipc import VisionIpcClient, VisionStreamType, VisionBuf
 from openpilot.common.swaglog import cloudlog
@@ -128,10 +128,15 @@ def main():
   assert vipc_client.is_connected()
   cloudlog.warning(f"connected with buffer size: {vipc_client.buffer_len}")
 
-  sm = SubMaster(["liveCalibration"])
+  sm = SubMaster(["liveCalibration", "carState"])
   pm = PubMaster(["driverStateV2"])
 
   calib = np.zeros(CALIB_LEN, dtype=np.float32)
+  # CLEARPILOT: cache last model output to serve while gear is in park —
+  # mirrors the same trick modeld uses. Skips DSP inference on the driver
+  # camera when the car is stationary; downstream dmonitoringd still gets
+  # a fresh publish each frame.
+  last_model_output = None
   # last = 0
 
   while True:
@@ -143,8 +148,13 @@ def main():
     if sm.updated["liveCalibration"]:
       calib[:] = np.array(sm["liveCalibration"].rpyCalib)
 
+    parked = sm["carState"].gearShifter == car.CarState.GearShifter.park
     t1 = time.perf_counter()
-    model_output, dsp_execution_time = model.run(buf, calib)
+    if parked and last_model_output is not None:
+      model_output, dsp_execution_time = last_model_output
+    else:
+      model_output, dsp_execution_time = model.run(buf, calib)
+      last_model_output = (model_output, dsp_execution_time)
     t2 = time.perf_counter()
 
     pm.send("driverStateV2", get_driverstate_packet(model_output, vipc_client.frame_id, vipc_client.timestamp_sof, t2 - t1, dsp_execution_time))