Drop mamba-ssm / causal-conv1d from requirements

HF Spaces build container has no nvcc, so mamba-ssm / causal-conv1d wheels
fail to compile at pip install time, aborting the whole build. The code
path already handles their absence gracefully — super_resolution lazy-loads
REUSEUpsampler and TTSServer._denoise_voice_ref catches the ImportError,
logs a one-time warning, and silently skips voice-ref denoise for the rest
of the session. Long-form chunking is unaffected.

requirements.txt

@@ -25,14 +25,14 @@ spaces>=0.30.0
 soundfile>=0.12.0
 resemble-perth @ git+https://github.com/resemble-ai/Perth.git@master
 
-# ── Optional: NVIDIA RE-USE speech enhancement (input-side voice-ref denoise) ─
-# RE-USE is applied to the uploaded voice reference before VAE encoding so the
-# model conditions on a clean speaker / style anchor (generated paralinguistic
-# events — laughs, breaths, sighs — stay untouched because the denoiser only
-# touches the reference). resampy is used for its pre-resample step; mamba_ssm
-# + causal-conv1d power the bi-Mamba kernels. The kernels have no pre-built
-# wheels for every CUDA toolkit — if installs fail, app.py logs a warning
-# once and silently skips the reference denoise for the rest of the session.
-resampy>=0.4.0
-mamba-ssm>=2.2.0 ; platform_system == "Linux"
-causal-conv1d>=1.4.0 ; platform_system == "Linux"
+# ── RE-USE deps intentionally omitted on HF Spaces ──────────────────────────
+# The mamba-ssm / causal-conv1d wheels need nvcc + matching CUDA toolkit at
+# install time, which HF Spaces' build container doesn't provide. Without
+# them, super_resolution.REUSEUpsampler raises ImportError on first use and
+# TTSServer._denoise_voice_ref logs a one-time warning + silently skips the
+# voice-ref denoise for the rest of the session. Long-form chunking + the
+# rest of the model are unaffected. To enable denoise_ref on the Space,
+# switch to a Dockerfile-based Space build with nvcc available, then add:
+#   resampy>=0.4.0
+#   mamba-ssm>=2.2.0 ; platform_system == "Linux"
+#   causal-conv1d>=1.4.0 ; platform_system == "Linux"