How should conflicting node-pack Python pins (torch, numpy, onnxruntime) be isolated in a single ComfyUI install?

Troubleshooting docs cite torch==2.4.1 vs torch>=2.4.2 conflicts, Impact-Pack pins numpy<2, controlnet_aux depends on CUDA-specific onnxruntime builds, and GGUF notes torch-version sensitivity on macOS. Sources describe symptoms but give no per-node isolation mechanism. Is per-pack venv or process isolation viable, and what does it cost in VRAM and latency? Why this is still open: No fetched source offers a tested isolation approach; the official page says resolving conflicts needs Python environment expertise and gives no commands.