fix: validate Qwen hybrid runtime placement

docs/conditional-testing.md

@@ -86,6 +86,27 @@ uv pip check --python .venv/bin/python
 .venv/bin/python scripts/run_conditional_gate.py bitsandbytes-runtime
 ```
 
+### Qwen3.5/Qwen3.8 hybrid runtime
+
+The Qwen3.5/Qwen3.8 Gated DeltaNet path requires both FLA and causal-conv1d.
+Install the optional extra only after selecting the CUDA build of PyTorch:
+
+```bash
+CUDA_HOME=/usr/local/cuda uv sync --locked --extra dev --extra qwen-hybrid
+.venv/bin/python -c 'import torch, fla, causal_conv1d; print(torch.__version__)'
+```
+
+`causal-conv1d` wheels are keyed to the Python, PyTorch, CUDA, platform, and C++
+ABI combination. Prefer an exact upstream wheel when one exists. A source build
+must use a `CUDA_HOME` whose major version matches `torch.version.cuda`; the
+system-default `nvcc` may point at a different toolkit. OBLITERATUS fails before
+weight allocation if either extension cannot be imported.
+
+The complete Qwen hybrid text model must fit one CUDA device with 15% free-memory
+headroom. Generic `device_map="auto"` layer sharding is deliberately disabled for
+this architecture because the recurrent-state execution path has not been
+validated across devices.
+
 Use the virtual environment interpreter directly after replacing Torch. A subsequent
 `uv run` or `uv sync` without the CUDA override may restore the portable CPU wheel
 from the lock. The agentic developer installer automates this ordering and performs