ci: verify ephemeral GPU runner end-to-end

[beveradb]

Summary

End-to-end verification for karaoke-gen PR #781 — the GPU image NVIDIA driver-load fix.

This PR is a single-line touch on audio_separator/separator/__init__.py whose only purpose is to trigger run-integration-tests.yaml's three self-hosted GPU jobs (ensemble-presets, core-models, stems-and-quality).

Expected behaviour

For each of the three jobs the ephemeral GHA runner dispatcher should:

1. Create a fresh GCE VM from the gha-runner-gpu image family (latest is gha-runner-gpu-20260518-035713).
2. Create it with Secure Boot OFF (per the enable_secure_boot=not family.has_gpu change in #781).
3. On first boot, gha-gpu-modprobe.service rebuilds nvidia.ko via DKMS against the running kernel (kernel skew between image-build and runtime is now handled), modprobes nvidia/nvidia_uvm/nvidia_drm, and verifies with nvidia-smi.
4. The "Verify GPU availability" step (nvidia-smi --query-gpu=driver_version,name,memory.total --format=csv,noheader) succeeds.
5. The actual integration tests run.

Smoke result before this PR

A smoke VM created by hand from gha-runner-gpu-20260518-035713 (SB off) booted successfully and nvidia-smi reported Tesla T4 / driver 595.71.05 / CUDA 13.2 as both root and as the runner user. Total boot-to-GPU-ready time: ~2 minutes.

Test plan

• All three self-hosted GPU jobs pass (ensemble-presets, core-models, stems-and-quality)
• Each job's "Verify GPU availability" step output shows the T4
• No "Failed to query NVIDIA devices" or "Key was rejected by service" entries in the runner logs

@coderabbitai ignore

🤖 Generated with Claude Code

[beveradb]

All 3 GPU integration jobs green — verifies karaoke-gen PR #781 end-to-end. Closing per the same convention as #286 (verification trigger, not a real change).