* Keep CTC alignment inside each item's input length.
A shorter row was scored through the padded frames, so a repeated label landed one frame early.
* Keep the CTC length mask on the score device.
Add an unequal-length batch test and a CPU-lengths test for an accelerator, and run tests.test_ctc_alignment in the container workflow.
* Make the unequal-length CTC batch fail on the old aligner.
The random batch agreed with a cropped item on the previous code. It now uses the five-frame example that moves when a padded frame is included.
