How I Turned Reference-Image Guesswork Into a Reproducible Test Harness

My image generations were failing in a very specific way: every output looked plausible, but I could not explain why one run kept the subject and another borrowed the camera angle from the wrong refer