The EPS floor for labels missing from top_logprobs is another useful sensitivity check. Temperature scaling is exactly a logit rescaling for a complete probability distribution, but assigning 1e-6 to unreported options introduces an approximation before the fit.

I would report how often the gold label was unreported at each step and repeat the fit with a few plausible floors. If T changes substantially, it is partly compensating for probability truncation rather than only model calibration. That would qualify the result without changing any of the selected answers.