fix: push diversity directive upstream into the visualizer

[meninoebom]

Summary

• flux-schnell-lora (the model we switched to in fix: route around stuck flux-schnell, bound latency, tame hands #73) has a stronger white-male-young-thin prior than the canonical pool. The style suffix's mild "varied skin tones" line gets ignored.
• The real lever is Claude, not the style suffix. Claude writes prose like "a figure leaning forward"; Flux fills the gap with its prior. If Claude instead writes "a middle-aged Filipina nurse leaning forward," Flux renders exactly that.
• Updates VISUALIZER_SYSTEM_PROMPT to require concrete demographic detail (ethnicity, age, gender, build) chosen to fit the theme, with explicit guidance not to default to a single prototype.
• Tightens STYLE_SUFFIX to back the directive: render figures exactly as described, don't fall back to white/male/young/thin.

Test plan

• 38/38 image tests pass
• Regenerate a few themes after deploy; eyeball that the rendered range feels varied across skin tones, ages, genders, and body types without feeling tokenistic
• If a single regenerate still skews narrow on one theme, that's expected — variety should be visible across themes, not within four candidates of the same prompt

🤖 Generated with Claude Code