Make NSFW flag a genuine visual check

The NSFW rule now tells the model to inspect the actual image for adult
content and flag it even when the bio text is innocuous, explicitly
overriding the text-first guidance.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

tempercheck/prompt.py

@@ -45,9 +45,14 @@ Score guide (0 = delightful, 100 = run):
   81-100 multiple strong warning signs (e.g. hostility + slurs/anger + conflict-pride)
 
 Two extra content flags, independent of temper — ALWAYS add the tag when present:
-- If the profile content looks NSFW / sexually explicit / adult (suggestive bio,
-  "18+", "spicy content", links to adult sites, explicit imagery), add the signal
-  "NSFW" AND mention the NSFW content in your rationale.
+- NSFW — inspect the ACTUAL IMAGE as carefully as the text, and flag adult
+  content from EITHER source. Add the signal "NSFW" if the picture shows nudity,
+  sexual or fetish content, underwear/lingerie as the focus, or other explicit or
+  adult visual material, OR if the text is sexually suggestive ("18+", "spicy
+  content", adult-site links). This visual check OVERRIDES the "text-first"
+  guidance above: flag NSFW based on the image even when the bio text is perfectly
+  innocuous. In your rationale, say whether the NSFW signal came from the image,
+  the text, or both.
 - If the profile mentions crypto (Bitcoin, Ethereum, NFTs, tokens, "$TICKER",
   "to the moon", web3, "diamond hands", trading shills), add the signal "crypto".
 Put any NSFW/crypto tags FIRST in the signals list so they are never dropped.