Ignore padding positions in the replaced-token detection loss

[gss10282025]

Summary

Replace RTD labels at attention_mask == 0 with -100 before the existing cross-entropy call, so padding contributes neither a negative-class target nor an element of the mean's denominator.

Valid-token labels, the RTD weight and the contrastive objective are unchanged. The reduction stays a mean.

Validation

At 33b29a38, one FP64 AdamW update on a small BERT, same batch padded to widths 8 and 16: the full-gradient difference drops from 1.04826645 to 8.90219e-13 normalized L2, the parameter-update difference from 0.79828865 to 4.46219e-12. A no-padding batch is unaffected. Reproduced on a second host.

The FP64 cast and the deterministic fixture belong to the reproducer, not to this patch. Earlier FP32 runs did not pass all validation checks; FP16 was not evaluated.

Issue links

Fixes #25

Checks on this branch

• Python syntax checks and git diff --check pass.
• The numerical results above come from the existing local reproductions; the repository training/CI suite has not been run on this branch.

Prepared with AI assistance; opened as a draft for code review.