A Global Color-Grading Bug in the Compositing Step

A batch of generated "warm kitchen counter" lifestyle images ships with products that appear slightly more orange/warm-toned than their actual color — confirmed by comparing against the seller's original phone photo and, in some cases, the same product's white-background image generated moments earlier in the same pipeline run. Product-region pixel-diff and SSIM scores for these images are elevated but still passed the fidelity gate, just barely under threshold.

1. What does "elevated but still passing" tell you about where the gate's threshold is currently set, and what's the risk of leaving it there?
2. Diagnose the likely root cause in the compositing pipeline based on the specific pattern described (only warm-toned templates affected, white-background images from the same run are fine).
3. Propose a fix, and explain why simply tightening the invariance threshold further is not, by itself, a complete solution.