Batched categorical kernel transform merges different categories #625

On current master af8b928, CategoricalParameter("x", ["a", "b", "c"]).kernel_transform(np.eye(3)) returns a 3x3 matrix of ones rather than identity. res[:, argmax(value, axis=1)] sets all selected columns in every row. For an RBF wrapped kernel this makes every pair of distinct categories identical (covariance 1 rather than exp(-1)), so a GP cannot distinguish them. Use paired row and category indices. A focused test checks mixed/repeated categories, square and rectangular wrapped-kernel reference values and the single-row control. The current test fails on untouched master. Searches of open/closed categorical kernel, kernel_transform and one-hot reports found no overlapping open PR; #601 concerns max conversion and #597 combined kernels.

AI assistance: Codex autonomously reproduced the issue and prepared the fix; no independent human review has occurred.