Rover clean

02 — A 4-motor rover controller (blind vs the loop)

The complex example: a real rover board — MCU (RP2040), IMU (MPU6050), four motor connectors, a battery entry, a USB-C data pair, a status LED, three decoupling caps — 14 nets, 13 components. Built twice, the way solidsight's engine block is:

• rover_blind.kicad_pcb is routed the way a fresh agent would with only the netlist and no way to SEE the copper: short diagonals laid between pads it cannot tell are in the way. It ships the defects that come from being blind.
• rover_clean.kicad_pcb is the same board taken to zero findings — routed by a real 2-layer maze router that reads each pad's actual position from pcbsight and keeps a clearance halo around everything it lays down.

python make_rover.py
pcbsight inspect rover_blind.kicad_pcb   # FAILED
pcbsight inspect rover_clean.kicad_pcb   # OK
pcbsight diff rover_blind.kicad_pcb rover_clean.kicad_pcb

Before (blind) vs after (the loop)

left: routed blind — diagonals crossing pads, findings circled in red, three motors and the IMU never reached. right: a clean 2-layer route, every net closed. Same components, same nets.

The blind board is not a strawman — it is exactly what routing without sight produces: 12 of 14 nets left open, 26 clearance violations, the USB pair swapped and unmatched, the battery rail necked to 0.25 mm. None of it is visible as "wrong" without measuring the copper.

The diff tells the whole repair

pcbsight diff rover_blind.kicad_pcb rover_clean.kicad_pcb
  tracks: 12 -> 113          vias: 0 -> 27
  net 'GND':    6 island(s) -> 1
  net 'USB_DP': 4 island(s) -> 1
  net 'USB_DM': 4 island(s) -> 1
  net 'M2_A' .. 'M4_B':      2 island(s) -> 1   (each motor pair reached)
  clearance findings: 26 -> 0
  ... every net-open and clearance finding GONE

Why the render looks like a board

Every component carries its silkscreen body, reference (U1, J2, D1...) and value (RP2040, MOTOR, LED...); the substrate is the real Edge.Cuts outline; gold is the front copper, dim gold the back. That is the point of the render overhaul — a PCB review has to look like a PCB, or the findings float in the void.

The clean side is genuinely routed, not decorated: _router.py is a small Lee/BFS maze router (two layers, vias, a clearance halo) so the "after" is a board that actually connects — the same standard the family holds everywhere: the fixed reference must really be fixed.