I have been checking the S2000 digital dash against a lit AP1 photograph, not against a feeling that the preview looks close enough. The comparison image puts that photograph beside the rendered face. The two panels are not the same width, so the seam is not the middle of the file.
The lit AP1 photograph used as the geometry reference. The Car Spy, CC BY 2.0, via Flickr.
Guessing the divider would shift every landmark. A speed readout that is in the right place on the module could fail, and a displaced one could pass, because the score was attached to the composite rather than the face. The capture step now writes the actual panel and module rectangles next to the image, with hashes of both the composite and the reference photograph.1
The measurement then looks only inside that recorded module. It finds the amber band, the red speed digits and the temperature and fuel bars, and compares those pixels with the documented AP1 lock. The older plate percentages stay in their original units. They are not divided by 100, and they are not the acceptance lock.
A blacked-out capture used to be easy to misread as a measurement of zero. Missing pixels now fail instead. Moving the speed digits sideways fails that landmark on its own. Changing the image after the metadata was written is rejected before any score is produced.
On the live capture, the band apex was 0.40 px from the lock, the speed height was 0.15 px under it, and the speed's right edge was 0.49 px inside it. Eight geometry checks passed, and the existing image-hash harness still passed. CI runs the capture and the check together.2
The follow-up was smaller, but it stopped the new test from breaking the others. The capture owns pygame's startup and shutdown. Running that inside the test process can tear down fonts that another render test still holds. The fixture now renders in a child process, and the measurement tests still pass when the display-fit tests run immediately afterwards.3
This is a geometry gate, not a claim that the face matches the car. The photograph is not a calibrated colour or font sample, and a still frame cannot show timing or motion. The cluster preview is still a bench mock. It is not a fitted dash, and these checks do not say the graphics are done.