test(accuracy): re-baseline HunyuanImage3 SSIM/PSNR for the vLLM rebase

HunyuanImage-3 (expert-parallel MoE) accuracy fails deterministically on
dev/vllm-align (SSIM 0.9490 vs 0.97, PSNR 26.21 vs 30 dB) while passing on
main's nightly. The align branch changes no hunyuan code; the drift comes
from the newer upstream vLLM (65a7b4628409, 585 commits past the v0.24.0
tag main tracks) shifting MoE kernel numerics. The image still renders
correctly and the mean/p99 pixel-diff checks pass — only the structural
metrics dip below a strict bar.

Per owner decision, re-baseline to the new stack: relax SSIM 0.97->0.94
and PSNR 30.0->25.0 (below the deterministic 0.9490 / 26.21, with margin).
The HunyuanImage CI job uploads no image artifact, so the reference PNG
cannot be auto-regenerated here; relaxing thresholds is the equivalent
re-baseline. main keeps the stricter bar.

Co-Authored-By: Claude <noreply@anthropic.com>

tests/e2e/accuracy/test_hunyuan_image3_pixel_accuracy.py

@@ -27,8 +27,15 @@ WIDTH = 1024
 PROMPT = "A brown and white dog is running on the grass."
 MEAN_THRESHOLD = 3e-2
 P99_THRESHOLD = 3e-1
-SSIM_THRESHOLD = 0.97
-PSNR_THRESHOLD = 30.0
+# NOTE: re-baselined for the vLLM rebase (dev/vllm-align). HunyuanImage-3 is an
+# expert-parallel MoE model; the newer upstream vLLM (65a7b4628409, 585 commits
+# past the v0.24.0 tag that main tracks) shifted MoE kernel numerics, so the
+# structural-similarity metrics against the committed baseline drift to a
+# deterministic SSIM 0.9490 / PSNR 26.21 dB (the mean/p99 pixel-diff checks above
+# still pass — the image renders correctly). Thresholds relaxed to accept this
+# drift as the new bar. On main (older vLLM) the stricter 0.97 / 30.0 hold.
+SSIM_THRESHOLD = 0.94
+PSNR_THRESHOLD = 25.0
 
 _REPO_ROOT = Path(__file__).resolve().parent.parent.parent.parent
 BASELINE_PATH = _REPO_ROOT / "tests" / "assets" / "hunyuan" / "hunyuan_baseline.png"