fix: push diversity directive upstream into the visualizer

app/images/providers.py

@@ -39,7 +39,12 @@
 - Describe what an observer would SEE. Start with a concrete noun phrase.
 - Translate abstract concepts into visual metaphors. "Autonomy" might become \
 "a single machine alone in a quiet glass room"; "attention" might become \
-"a figure leaning forward over a steaming cup."
+"a middle-aged Filipina nurse leaning forward over a steaming cup."
+- When the scene includes a person, name them with specific demographic detail — \
+ethnicity, age, gender, build, distinguishing features — chosen to fit the theme. \
+Don't default to one prototype (white, male, young, thin); across themes the \
+rendered range should feel like the world. Vary skin tones, ages, gender \
+presentations, and body types naturally, as the theme allows.
 - Do NOT repeat the theme's phrasing verbatim. Interpret, don't echo.
 - Do NOT use meta-language ("a painting showing X", "an image of Y"). Just the scene.
 - Do NOT include any style or medium notes — those are added downstream.

app/images/service.py

@@ -13,7 +13,9 @@
     "Rendered as a flat oil painting in a limited three-color palette, "
     "figurative and confident, contemporary museum-quality painting, "
     "tone balanced between gravity and play. "
-    "If human figures are included, represent varied skin tones, ages, and body types. "
+    "Render any human figure exactly as the scene describes; do not default to a "
+    "white or male or young or thin figure when the scene calls for someone "
+    "different. "
     "When hands are visible, they rest calmly at the sides, are folded, or hold "
     "a single object — never reaching, gesturing, or pointing."
 )