[chu-jiaming]

What

• Add camera_correction.png: T5 training-camera centers before and after
  Naive + warm-up, with errors drawn ×10
• Add the figure to both READMEs (Results §1) and document --gt-transforms

Why

The README showed the drift in image space only; this shows the geometric
side: pose optimization pulls cameras from 0.16 to 0.025 units of error.

How to verify

• The T5 perturbation is regenerated from GT with seed 42; the script checks it
  against the archived initial errors and aborts on mismatch
• CPU tests pass