bench: multi-reference ref2v support (1-8 ref images, match/max sizing)

- refheavy suite: ref2v saturation with 1/4/8 refs, the priority modality
- family suite: reduced grid for int8/bf16 family comparisons
- 8 distinct 1344x768 reference scenes generated into /ComfyUI/input

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

scripts/pod/make-test-images.py

@@ -114,8 +114,35 @@ def bench_ref():
     return img
 
 
+# Distinct palettes for the multi-reference ref2v benchmark set. Each image is
+# a different scene so 4-8 refs exercise genuinely different reference context.
+REF_PALETTES = [
+    ((200, 120, 60), (240, 200, 140)),   # amber sunset
+    ((20, 30, 80), (90, 140, 200)),      # blue dusk
+    ((10, 60, 50), (120, 200, 160)),     # teal dawn
+    ((80, 20, 90), (200, 120, 220)),     # violet night
+    ((160, 40, 30), (250, 150, 90)),     # red dusk
+    ((30, 30, 30), (140, 140, 150)),     # overcast grey
+    ((10, 20, 40), (40, 90, 160)),       # deep night blue
+    ((190, 160, 40), (250, 240, 170)),   # golden noon
+]
+
+
+def bench_ref_set():
+    """8 distinct city scenes, bench_ref_01..08.png (1344x768)."""
+    for i, (top, bottom) in enumerate(REF_PALETTES, start=1):
+        img = city_backdrop(1344, 768, top, bottom, seed=100 + i)
+        d = ImageDraw.Draw(img)
+        # distinct celestial marker per scene so refs are visually distinct
+        x = 120 + i * 130
+        d.ellipse([x, 80, x + 90, 170],
+                  fill=(255, 240, 200) if i % 2 else (240, 240, 255))
+        img.save(f"{OUT}/bench_ref_{i:02d}.png")
+
+
 if __name__ == "__main__":
     superboy().save(f"{OUT}/red_superboy_on_city_roof.png")
     mecha_dragon().save(f"{OUT}/mecha_dragon_lightning.png")
     bench_ref().save(f"{OUT}/bench_ref_1344x768.png")
-    print("wrote 3 images to", OUT)
+    bench_ref_set()
+    print("wrote 11 images to", OUT)