onnxruntime 1.23.2, identical graph bytes, identical input bytes,
intra/inter_op_num_threads=1, ORT_SEQUENTIAL, ORT_DISABLE_ALL.

elements                49152
bit-identical           NO
elements differing      40746  (82.9%)
max |delta|             1.907e-06
mean |delta|            2.538e-07
max relative delta      7.726e-03
max ULPs                81920

fux sorts on round(score, 9); a score of order 1 drifts ~1.9e-06,
which is ~2000x the rounding. One encoder block. A 6-layer model compounds it.
