_normalize_attn_mask_for_sdpa passed tokenizer-produced int64 0/1 masks
straight to F.scaled_dot_product_attention, which rejects long dtypes — the
HYWorld pipeline crashed on exactly the SDPA-fallback machines the fastcheck
stabilization targeted; coerce non-bool/non-float masks to bool. Lift the
[batch, key_len] -> [batch, 1, 1, key_len] reinterpretation out of the impl
into SDPAMetadataBuilder.build, where that semantic is known, so 2D masks
reaching the impl keep torch's documented [query_len, key_len] broadcast
meaning. Reuse the identical metadata for HYWorld's prope attention instead
of rebuilding it, and note that the metadata does not choose the executing
kernel.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
