diff --git a/README.md b/README.md
index dfcc58e..84597f3 100644
--- a/README.md
+++ b/README.md
@@ -73,7 +73,7 @@ does all inference over HTTP).
 
 ```bash
 # 1. Build everything: venv (auto-selects AMD/NVIDIA/CPU), web UI, memory DB,
-#    system packages and the Ollama binary.
+#    system packages, Ollama binary and the XFCE desktop wiring.
 ./install.sh
 
 # 2. Launch (memory :8001, backend :8000 — backend also serves the built UI)
diff --git a/bin/gen-nvidia-reqs.py b/bin/gen-nvidia-reqs.py
index 56d1f6d..8e1a4df 100644
--- a/bin/gen-nvidia-reqs.py
+++ b/bin/gen-nvidia-reqs.py
@@ -43,6 +43,9 @@ def main():
     suffix = wheel_suffix(major, minor)
     print(f"PyTorch wheel: {suffix}")
 
+    # newline="\n": write_text() otherwise translates to os.linesep, so running
+    # this on the Windows box produced a CRLF requirements file that then showed
+    # up as a whole-file diff every time it was published.
     NVIDIA_REQS.write_text(f"""\
 # --- Force NVIDIA/CUDA Priority ---
 --index-url https://download.pytorch.org/whl/{suffix}
@@ -54,7 +57,7 @@ def main():
 torch
 torchaudio
 torchvision
-""")
+""", newline="\n")
 
     print(f"\nWritten: {NVIDIA_REQS}")
     print("Run 'ncp backup' to push it to the router.")