Force Stop (cem_path) was missing a tracking_lead guard, causing it to activate behind stopped leads at red lights and stop signs — stopping the car far short of the lead instead of letting ACC handle following distance. Added and not self.starpilot_planner.tracking_lead to cem_path, matching the guard already present on dash_path.
F1 — Toggle cache (starpilot_annotated_camera.h + .cc): Added 25 typed cached members; all 32 starpilot_toggles.value() calls in paint code replaced with the cached versions. All reads happen once at the top of updateState() per frame. The only surviving raw call is in mousePressEvent (intentional — touch is rare, not paint-path).
F2 — HUD gradient hoist (selfdrive/ui/qt/onroad/hud.cc): Header gradient is now a static const built once via lambda, no allocation per frame.
Build, flash, and verify that toggling compass, simple_mode, radar_tracks etc. from settings still take effect live without a restart.
D2 — update() in starpilot_onroad.cc now only fires when at least one overlay (showBlindspot, showSignal, showSteering, showFPS) is active. ~30–60 wasted repaints/sec eliminated when all overlays are off.
E4 — Holiday theme check reordered so current_holiday_theme != "stock" (cheapest) is first. Default is stock, so the rest of the chain never evaluates on a normal non-holiday drive.
E7 — paintSteeringTorqueBorder no longer rebuilds the QLinearGradient and computes lighter(120)/setAlpha/5 setColorAt calls every frame. The gradient is now built once in resizeEvent and reused. The per-frame work is now just the two geometry calculations and fillRect.
E5/E6/E8 — Audited clean: all remaining params_memory.put calls are event-driven (button presses, random events), no hot-path logging exists in the UI code, and all QTimers are event-driven or single-shot offroad dialogs.
Wave A (pure refactors):
A2: nnff toggle moved to front of NNFF check — kills ~100 IPC reads/sec when NNFF disabled
A4: 3 filters now use update_alpha() instead of being recreated — ~60 allocs/sec eliminated
Wave B (write reduction):
B1: CEStatus only writes on state transitions (+ E3 folded in: extracted ce_write_value local at both sites). Reset path added in the manual-override branch. ~19 writes/sec saved during cruise
B2: GPS position only persists when compass is enabled AND bearing shifts >0.5° (+ E1 folded in: json.dumps inside the gate). ~95% of writes eliminated on highway
Wave C (polling throttles):
C1: Onscreen distance button reads at 10 Hz instead of 100 Hz — ~90 IPC reads/sec eliminated. Wheel button unaffected
C2: SpeedLimitAccepted IPC read throttled to 5 Hz via existing _slc_adopt_counter. Wheel accel still instant
C3: Lane width numpy calls at 5 Hz instead of 20 Hz — ~300 allocs/sec eliminated. Counter resets when speed drops below threshold so values are fresh on re-entry
firestar5683