feat(gtx-960m): leave the nvidia driver unloaded so the dGPU stops drawing idle power

[ExtraToast]

Stacked on #763.

Problem

The GTX 960M on enschede-gtx-960m-1 has never powered down once — not in 29.5 days of uptime:

runtime_suspended_time = 0 ms
runtime_active_time    = 2552506816 ms   (29.5 days)
power/control = on      runtime_status = active

Meanwhile the card is doing nothing: P8, 44C, 0% utilisation, 3 MiB of 2048 MiB used, no processes.

Why it cannot be fixed with configuration

The driver states the limit directly:

Runtime D3 status:          Disabled by default
GPU Hardware Support:
 Video Memory Self Refresh: Not Supported
 Video Memory Off:          Not Supported

Per the NVIDIA driver documentation, runtime D3 requires a Turing or newer GPU, and NVreg_DynamicPowerManagement: 3 means fine-grained control on Ampere+ notebooks and disabled everywhere else. This GPU is Maxwell 2.0 (GM107).

So the card idles at full power for as long as any driver is bound to it, regardless of which processes hold /dev/nvidia*. Stopping wolf or the device plugin does not change this — that was the wrong diagnosis. There is no supported way to suspend a Maxwell dGPU under the proprietary driver.

Change

The driver is no longer loaded on this host:

• profiles/gpu-nvidia.nix and modules/services/game-streaming.nix dropped from the host imports (the latter forced boot.kernelModules = [ "nvidia" ... ]).
• nouveau and the four nvidia* modules blacklisted.
• A udev rule enables PCI runtime PM on the NVIDIA display controller so the kernel can take the now-unbound device out of D0.

Inventory follows: the node's gpus entry and its nvidia / game-streaming capabilities are removed, host_native no longer lists game-streaming for it, and jellyfin loses temporary_gpu_model: gtx960m.

GPU work stays on enschede-t1000-1 (transcode) and enschede-rx7900xtx-1 (render-compute), which keep their capabilities. nvidia-device-plugin and dcgm-exporter select on capability-nvidia, so both simply stop scheduling here and continue on t1000.

Dead code

Once this host stopped importing it, game-streaming.nix had no importer at all — enschede-rx7900xtx-1 uses the self-contained game-streaming-amd.nix. The module is deleted. Its wolf app-catalog assertions move onto the AMD module, which is the host that actually serves game streaming, and a new test pins the no-GPU invariant for this host.

Consequence that needs a follow-up decision

wolfmanager selects personal-stack/capability-game-streaming. After this change only enschede-rx7900xtx-1 carries that label, and that node has been NotReady for 90 days — so wolfmanager will be unschedulable and the wolf public host will not serve until the AMD node is back. Options are to bring that node up or retire the wolf ingress; retiring a public service is out of scope here.

Validation

• cd platform/tests && node --test — 58/58 pass (the gtx test is split into three: the no-GPU invariant, the AMD game-streaming host, and continued gpu-nvidia coverage for t1000).
• All five render scripts run clean, no rendered-output drift.
• kubectl kustomize platform/cluster/flux/clusters/production builds.
• nix flake check was not run locally — nix is not installed on this workstation. Platform Validate runs it in CI and must be green before merge.
• Whether the unbound card actually reaches a low-power state is verifiable only after the rebuild, by re-reading runtime_suspended_time on the host.