refuse local media jobs when the vLLM container is holding the GPU

[atomantic]

Summary

• A serving vLLM Qwen3.8-27B container holds ~23 GB of a 24 GB RTX 3090, and unloadResidentModels() cannot see it — it evicts loopback Ollama / LM Studio only. Every GPU-heavy local job proceeded, allocated, and died inside its model load with an OOM naming neither vLLM nor a remedy.
• New detectGpuBlockers() in server/lib/localMemory.js probes an enabled vllmBacked provider's local endpoint and surfaces the result as an additive blockers array on the prepareLocalMemory() report.
• All four GPU-heavy callers (local image, video, image-to-3D, LoRA training) now fail fast on a non-empty blockers, with prose naming the container, its project directory, and the docker compose --profile single stop that frees the card.
• Detect and refuse, never auto-stop — nothing would restart the container, an attached agent session dies with it, and a cold start is ~5–7 minutes. This mirrors PortOS's existing refusal to start it.
• No latency on the common path: the probe is skipped entirely with no enabled vllmBacked provider and on any non-CUDA host. Per the sentinel rule, a probe that fails, times out, or throws means "not serving" → the job proceeds; only an endpoint that actually answers blocks (including a 401/403, which is a container up behind VLLM_API_KEY).
• Nothing is unloaded while blocked — evicting the user's resident LLM for a run that is about to be refused helps nobody.
• Rolled in: the image-generation pre-wiring failure path now unwinds its jobs entry and stepwise temp dir the way the busy branch already does, instead of only releasing the claim.

Test plan

• cd server && npm test — 1565 files, 32,773 tests green.
• New server/lib/localMemory.test.js (17 cases): no probe with no enabled vllmBacked provider (and no CUDA probe either), no probe when it is disabled, no probe on a non-CUDA host, still probes on an unknown CUDA verdict, blocker when the endpoint answers, blocker on a 401, no blocker when the probe fails / times out / throws, no blocker for a vLLM on another machine, one probe per endpoint preferring the credentialed record, no blocker when the provider store cannot be read, and the message naming the container / project dir / stop command.
• Caller-level: server/services/videoGen/local.test.js asserts a blocked render rejects with the stop-command prose, never reaches spawnDetached, and releases the heavy claim.
• Existing prepareLocalMemory mocks updated to carry blockers, so a schema drift fails rather than silently passing.

Closes #4766