Fixes #16247.

Problem
In DuplexSTTDataset._apply_early_interruption_augmentation (nemo/collections/speechlm2/data/duplex_stt_dataset.py), early interruption is simulated by placing eos_id at new_eos_pos, shifting the tail if applicable, and then overwriting trailing frames with pad_id:

target_tokens[batch_idx, new_eos_pos] = eos_id
seq_len = target_tokens.shape[1]
cont_start_pos = original_eos_pos + overlap_tokens
tail_length = seq_len - (cont_start_pos + 1)
if tail_length > 0:
    target_tokens[batch_idx, new_eos_pos + 1 : new_eos_pos + 1 + tail_length] = target_tokens[
        batch_idx, cont_start_pos + 1 : cont_start_pos + 1 + tail_length
    ].clone()
target_tokens[batch_idx, -frames_to_remove:] = pad_id

When original_eos_pos sits near the end of the sequence such that cont_start_pos >= seq_len - 1 (no tail-shift runs), frames_to_remove = original_eos_pos - cutoff. Because new_eos_pos = cutoff, the pad-fill range [-frames_to_remove:] starts at or before new_eos_pos, immediately overwriting the relocated eos_id with pad_id. The augmented turn ends up with no EOS token at all.

Fix
Calculate the pad start boundary as pad_start = new_eos_pos + 1 + max(0, tail_length) and set target_tokens[batch_idx, pad_start:] = pad_id, and ensure target_tokens[batch_idx, new_eos_pos] = eos_id is placed so that the relocated EOS token is strictly preserved.

Testing

• Added unit test in tests/collections/speechlm2/test_duplex_early_interruption_eos.py exercising the boundary condition (seq_len=25, original_eos_pos=24, overlap_tokens=5) and asserting that the relocated EOS token is preserved.
• Verified test passes on CPU.