[inureyes]

Summary

The CUDA promotion overlay (patches-cuda/dtype.cpp, kept unchanged) resolves bf16 with f32 to bf16, so stored bf16 tensors demoted the f32 EAR-TTS backbone stream on CUDA. This fixes all four sites inside the VoiceChat path:

• bos_emb and audio_prompt_projection_W are promoted to f32 at load (promotes_to_f32).
• Gemma3Backbone::widen_norms_to_f32 (called only by the VoiceChat loader) rebuilds each backbone norm (4 block norms, q_norm, k_norm, final norm) so it holds the stored-dtype 1 + w widened to f32. Exact for half-precision weights (proof in the doc comment); the shared GemmaRMSNorm in layers.rs is unchanged.
• The gated fusion casts its bf16 text branch to the audio dtype. The issue did not list this site, but multiply(f32, bf16) there also returned bf16 on CUDA.
• null_emb stays bf16, narrowing the issue's proposal: it is concatenated with the bf16 subword condition, so promoting it would change the text_proj input on non-CUDA builds and break identity. Its existing "not promoted" assertion is kept; bos_emb and audio_prompt_projection_W move to the promoted list.

Validation

• New tts_dtype_tests::backbone_stream_is_f32_with_bf16_weights (all-bf16 tiny model): stream dtype after the projection and bos_emb assembly, after the fusion, after the final norm, and on a step. On 33c4505 with --features cuda it fails at the first assertion (left: 12 bf16, right: 10 f32).
• New widened_norms_hold_the_stored_one_plus_w_in_f32: byte equality of every widened 1 + w against the cast stored one, including w = -1, -1 +/- 1 ulp, and large values, plus an f32 stream through a norm and the backbone.
• GB10 --release --features cuda, --test-threads=1 under gpu-lock: models::nemotron_voicechat 28/28, models::gemma3_backbone 6/6, audio::f32_weights 4/4. clippy --lib --tests -D warnings and cargo fmt --check clean.
• Non-CUDA identity is argued from the code, not measured (CPU-only Linux build is blocked by fix: CPU-only Linux build fails to link kv_inplace_write #2108): each change is the exact bf16 to f32 astype that upstream promotion already inserts inside matmul, add, multiply and fast::rms_norm, so the kernels read identical values.
• No real VoiceChat checkpoint was used; none is available on GB10.

Closes #2109