T2: force canary failure via bogus image tag

Override gptoss-fp4-h100-vllm image to a nonexistent vLLM tag so the
canary's docker pull fails. Expected workflow behavior under the new
tighter gate:
  canary-sweep.result == 'failure'
  → all 8 fan-out jobs skip (gate: success || skipped)

Will be reverted after T2 verifies.

Author: Oseltamivir

.github/configs/nvidia-master.yaml

@@ -4492,7 +4492,7 @@ minimaxm2.5-fp4-b300-vllm:
       - { tp: 8, conc-start: 4, conc-end: 4 }
 
 gptoss-fp4-h100-vllm:
-  image: vllm/vllm-openai:v0.21.0
+  image: vllm/vllm-openai:nonexistent-tag-T2-canary-failure-test
   model: openai/gpt-oss-120b
   model-prefix: gptoss
   runner: h100

perf-changelog.yaml

@@ -3133,5 +3133,5 @@
 - config-keys:
     - gptoss-fp4-h100-vllm
   description:
-    - "[T1 / PR #1503 canary-gate test] No-op trigger to exercise canary success + dedup'd fan-out on the existing gptoss h100 vLLM config (h100 chosen for shorter CI queue). Run-sweep PR also tightens cancel-gate to `success || skipped` and adds new `non-canary-full-sweep-enabled` label (full sweep without canary)."
+    - "[T2 / PR #1503 canary-gate test] Image override to nonexistent vLLM tag forces canary image-pull to fail. Expected: canary-sweep.result == 'failure', all fan-out jobs (single-node/multi-node/agentic/evals) skipped via the new `success || skipped` gate. T1 already verified canary success path on the same config."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1503